Content uploaded by Emine Bozkuş
Author content
All content in this area was uploaded by Emine Bozkuş on Dec 23, 2021
Content may be subject to copyright.
450
TAM METİN SÖZEL SUNUMLAR
EVALUATION OF OCCUPATIONAL ACCIDENTS WITH ARTIFICIAL NEURAL
NETWORKS IN OCCUPATIONAL HEALTH AND SAFETY MANAGEMENT SYSTEMS
İrem SAHMUTOGLU1, Fatma Talya TEMIZCERI 2, Emine BOZKUŞ3
1 Bartın University, Vocational School of Higher Education, Department of Administration and
Organization, Bartın / Turkey
2 İstanbul Bilgi University, School of Applied Science, Department of Logistics Management,
Istanbul / Turkey
3 Yildiz Technical University, Faculty of Mechanical, Department of Industrial Engineering,
Istanbul / Turkey
Abstract: Taking continuous and efficient precautions to possible accidents that may occur in all indus-
trial organizations is the main requirement for the health and safety of employees. For this reason, Oc-
cupational Health and Safety Management Systems are practiced to improve systematically every type
of hazard, risk, and risk assessment process. Organizations are expected to use quantitative risk assess-
ment methodologies to minimize possible accidents at workplaces. In our study, Artificial Neural Net-
works (ANN) approach was designed to support occupational health and safety management systems as
a risk assessment by a quantitative method. Within the scope of the study, the number of incapacities to
work forecasting models was developed for the employees who have compulsory insurance by using
artificial neural networks. The number of employees with compulsory insurance from 1970-2018, the
number of workplaces, work accidents, and the number of people who died consequently because of
work accidents, and permanent disability data were the input data for this study. Multi-Layer Perceptron
(MLP), which is one of the ANN topologies, is developed, and the performance was tested according to
its (Mean Squared Error) MSE and Root Mean Squared Error (RMSE) values. It has been observed that
the MLP network produces better results with the best MSE value as 0.2250 and the best RMSE value
as 0.1125.
Keywords: Occupational Accidents, Occupational Health and Safety, Risk Assessment, Artificial Neural
Network
INTRODUCTION and CONCEPTUAL FRAMEWORK
Occupational accidents are a global problem in occupational health and safety systems. Despite the cur-
rent regulations to prevent occupational accidents in the world, approximately 2.3 million people die
hereby-occupational accidents every year (ILO), which makes more than 6000 deaths per day. World-
wide, 340 million occupational accidents and 160 million people have occupational diseases annually25.
ILO updates these data periodically, and the statistics show that accidents and illnesses are unfortunately
increasing.
Accident rates are high in developing countries. Turkey ranks third in the world after Algeria and El
Salvador for accidents involving a death at work and is located at the top of the list in Europe for occu-
pational accidents. The rate of occupational accidents involving deaths in Turkey is obtained as 20.5
employees per 100,000 people, while this rate is 2 per 100,000 workers in countries such as Norway,
25 https://www.ilo.org/global/topics/safety-and-health-at-work/lang--en/index.htm
TAM METİN SÖZEL SUNUMLAR
451
Sweden, and Denmark
26
. Only in 2018, 431276 occupational accidents happened, and 1541 people died
because of these accidents, while 3773 people were permanently unable to work
27
.
The concept of occupational accidents has different definitions in the literature. According to the ILO,
an occupational accident is defined as an unexpected and undesirable incident during an economic ac-
tivity that causes injury or loss of one or more workers, while Occupational Health and Safety Act No.
6331 defines it as an event that causes death or apologizes for body integrity spiritually or physically.
OHS experts and practitioners have continuously been looking for methods to analyze the relationship
between possible occupational diseases or disorders on workers and to predict the outcome of any vari-
able. The most popular tools among these methods are statistical analysis and logistic regression (LR).
Studies have been carried out for prediction models with artificial neural networks (ANN) and this
method has also been used for work accident prediction models for the last 25 years. There are different
studies on this subject in the literature.
The first study published by Karwowski et al. in the book titled Neural Network in Occupational Health
and Safety classified the jobs based on the Artificial Neural Networks method for low back pain syn-
dromes. Back / low back pain is a common occupational disease in the literature. This study was iden-
tified as a pilot study in Kentucky and the jobs that pose a potential risk for back pain were classified by
ANN. This study aimed to guide the employer especially for hazard analyses and injury measures for
cargo handling jobs. By the developed method, operations were successfully classified as low and high-
risk groups for heavy lifting and handling tasks (Karwowski et al., 1994).
Kaçar estimated the load coming to the articulation points by using the backpropagation network from
artificial neural networks in the carrier construction machines used in the industry. He also ran the pro-
posed model with simulation and concluded that the neural network values and simulation values are
compatible
28
.
Moayed and Shell aimed to show that ANN models execute better than LR models. The data set drawn
in this research was collected from construction workers using the Job Compliance survey. The dataset
was prepared as both input and output (result) variables. LR models and ANN models were generated
using the identical data set and the performance of all models held a candle to using the log-likelihood
ratio. This study demonstrated that ANN models have significantly better performance than LR models
(Mooyed&Shell, 2010: 132-142).
Beriha et al., proposed a factor analysis to evaluate the perceptions of safety related to Occupational
Health and Safety (OHS) norms among the labour force in Indian industries and to develop a tool for
evaluating these OSH norms in the three major industry sectors with a wide questionnaire. Outputs are
defined as injury level and material damage. To explain the relationship between input parameters and
outputs, they have adopted a neural network approach to conduct sensitivity analysis. They conducted
an explanatory factor analysis and interpreted the responses given to the questionnaire. Due to the neural
26
https://ilostat.ilo.org/data/
27
http://www.sgk.gov.tr/wps/portal/sgk/tr/kurumsal/istatistik/sgk_istatistik_yilliklari
28
http://www.emo.org.tr/ekler/7dd0be2efb27309_ek.pdf
TAM METİN SÖZEL SUNUMLAR
452
networks that can easily model the behaviour and security detection mechanisms of workers, a compar-
ative analysis was performed based on the ANN approach that was proposed for the current applications
in the Indian industry (Beriha et al., 2012: 180-200).
Goh and Chua proposed an ANN approach to accident data from the Singapore construction industry.
They conducted a case study to understand the relationship between the OSHMS elements and safety
performance. The findings implied that the learning performance of the ANN from incidents reduces
the severity and likelihood of accidents at the construction sites in Singapore. The authors also showed
that the ANN approach was successful to obtain meaningful information on how to improve safety per-
formance (Goh&Chua, 2012: 460-470).
Ceylan and Avan estimated the work accident, permanent incapacity to work, and the number of deaths
with three different scenarios until 2025 by using artificial neural networks. They used the number of
employees with insurance, workplace, work accident, dead, and incapacity to work as parameters be-
tween the years 1970-2010. They applied 2-5-1 neural network architecture and set the training set be-
tween 1970 and 2004 to estimate values between 2005 and 2010. They evaluate the model performances
using Mean Absolute Percentage Errors (MAPE), Mean Absolute Errors (MAE), and Root Mean Square
Errors (RMSE). Additionally, they compared the outputs with real values and concluded that the model
was valid for this purpose (Ceylan&Avan, 2013: 46-54).
Ceylan extended the study which was published previously with three different scenarios until 2025 to
Turkey by using artificial neural network models with an extended data set between 1970 and 2012. The
author concluded that the best network architecture type by using sigmoid and linear functions for dif-
ferent layers and used the feed-forward backpropagation algorithm with the training set between 1970
and 1999, to estimate the values between 2000 and 2012. The performances of models with ANN logic
were evaluated using mean absolute percent errors (MAPE), average absolute errors (MAE) and root
mean square errors (RMSE) (Ceylan, 2014: 1-10).
Especially occupational injuries in large-scale workplaces and production sites are of great importance
as a health problem. The most basic way to protect the health of the workforce at workplaces and to
improve conditions is the analysis and modelling of health-threatening factors. Accordingly, Moham-
madfam et al. aimed to determine the factors that threaten health in their studies and explained the se-
verity of occupational injuries by ANN modelling to provide a model to estimate the severity of occu-
pational injuries. 10-year data covering the years 2005-2014 were taken from 10 major construction
sectors. 960 work accidents were analysed based on coarse cluster theory and feature weighting with
artificial neural networks (ANNs) and modelled using RSES and MATLAB. Findings showed that the
severity of injuries resulted in various factors such as individual, organizational, health, and safety train-
ing and risk management factors of accident severity (Mohammadfam et al., 2015: 1515-1522).
Alaeddinoğlu et al. developed a new system to perform more effective and adequate risk analysis and
assessment by using artificial neural networks (ANN) with the accident data from 2009-2014. As a result
of the tests, the system was found to be feasible (Alaeddinoğlu et al., 2015: 275-291).
Dizdar and Koçar developed different models to interpret the probability of work accidents in institu-
tions where there occur critical OSH hazards, considering ANN. The coefficients of the performance
determinants (R2) and the Root Mean Square Error were evaluated.
TAM METİN SÖZEL SUNUMLAR
453
Abubakar et al. proposed a model for the interaction between organizational safety concepts and acci-
dents in the workplace. An artificial neural network was designed using survey data from 306 metal
casting industry employees selected from Central Anatolia. They observed that the “internal security
climate” phenomenon reduced workplace injuries. The authors discussed the theoretical and practical
implications of reduced workplace injuries in the Anatolian metal casting industry (Abubakar et al.
2018: 1-11).
Ayanoğlu and Kurt built a data set by analyzing the accidents in the metal sector workplace. After ap-
plying multivariate data analysis methods on the data set, they reduced the variables of the data set. They
determined that the algorithm produced the best estimation by artificial neural networks. Besides, they
developed an accident prediction model by using double-layer feed-forward ANN and evaluated possi-
ble accident risks of the metal industry workplace sample (Ayanoğlu&Kurt, 2019: 78-87).
Ayhan and Tokdemir examined the 829 accidents with a sharp object from 55 different construction
sites. Using the Analytical Hierarchy Process (AHP) method, they calculated the effect of accident var-
iables on the type of accident. They developed an ANN model for prediction. They partitioned the data
set for training and testing. The best model reached 78% accuracy. Their proposed work allows profes-
sionals to anticipate the potential risks and scenarios of “Sharp Object Contact” accidents and to under-
stand the trigger factors in detail both before and during construction (Ayhan&Tokdemir, 2019: 323-
334).
Qiao et al. integrated ANN to identify and classify the human factors involved in maritime accidents.
(Qiao et al., 2020)
Recal and Demirel classified the severity of occupational accidents with different Machine Learning
methods such Support Vector Machine, Multinomial Logistic Regression, Neural Network and etc. (Re-
cal and Demirel, 2021)
Kho et al. classified the severity of road accidents with different Machine Learning methods such as
Support Vector Machine, Random Forest and Neural Network. (Kho et al., 2021)
This paper provides a parametric study for the network structure that gives better results by making
experiments in MLP network structure on the Social Security Institution (SSI) data of the years between
1970-2018. This structure is tested with different parameters. Test results showed that the MLP network
structure provided reasonable results. The contribution of this study to the literature is that the work
accident data has not been tried before a parametric study for the MLP network with the current work
accident data.
MATERIAL and METHOD
Artificial neural network models have been developed for the output to be obtained with the data from
the study, in addition to the information obtained from the Social Security Institution and Turkish Cham-
ber of Engineers and Architects Association's annual reports from 1970-2018 (Ceylan, 2014). Artificial
Neural Networks (ANN) is one of the artificial intelligence techniques that model using the working
structure of the human brain. ANNs are very effective in model estimation. Through trial and error, it
establishes relationships between data, learns and in this way can manage unknown processes. In this
study, MLP network structure was used for the training and testing phase of SSI data.
TAM METİN SÖZEL SUNUMLAR
454
ANNs collect information about examples, generalize, and produce outputs using the information they
learn. Due to these learning and generalizing features; artificial neural networks show the ability to
successfully solve complex problems (Elmas, 2018). According to another definition, ANNs; infor-
mation processing structures inspired by the human brain are interconnected by weighted links and each
consists of its memory processing elements; in other words, computer programs that imitate biological
neural networks (Elmas, 2013). One of the most distinguishing features of ANNs is the ability to learn.
Learning is defined as the calculation of the link weights that can ensure that the structure between the
examples available can behave well. ANNs store information obtained during learning as link weights
between nerve cells. These weight values contain the necessary information for ANN to process data
successfully (Şen, 2004).
Data Collection
The data from 1970 to 2006 presented in Table 1 were obtained from the study of Ceylan and Avam,
and the data from 2007 to 2018 were obtained from the reports published by the Social Security Institu-
tion on its website. This data set consists of the year, the number of compulsory SSI members, work-
places, accidents, deaths, and workers who are incapacity to work. There are 5 inputs and 1 output for
the MLP network.
Table 1. SSI data between 1970-2018
Year
# of In-
sured Em-
ployees
# of Busi-
ness
# of Work
Accidents
#of
Deaths
# of Incapacity
to Work
1970
1313500
109391
144483
679
284
1971
1404816
154812
148822
583
2574
1972
1525012
174344
160585
682
2359
1973
1649079
184427
176993
822
2372
1974
1799998
195929
180375
983
2643
1975
1823338
205441
182601
855
2560
1976
2017875
216941
196341
947
2659
1977
2191251
229198
199961
1135
3123
1978
2206056
231130
193998
975
2841
1979
2152411
239225
186089
1050
2053
1980
2204807
241580
159600
1014
2406
1981
2228439
259589
165101
938
2300
1982
2264788
273226
147118
831
1881
1983
2327245
281627
144483
1070
2592
1984
2439016
294284
148822
885
2453
1985
2607865
326996
160585
877
2549
1986
2815230
365514
176993
1108
2282
1987
2878925
387452
180375
838
2483
1988
3140071
451662
182601
1163
2170
1989
3271013
474318
196341
1150
2394
1990
3446502
514390
199961
1292
2778
1991
3598315
536098
193998
1189
3334
1992
3796702
559184
186089
1583
3044
TAM METİN SÖZEL SUNUMLAR
455
1993
3976202
610129
159600
1064
3522
1994
4202616
691023
165101
1034
2791
1995
4410744
724427
147118
798
2188
1996
4624330
759342
145296
1296
2249
1997
4830056
781911
152650
1282
3445
1998
5299533
813010
148027
1094
2677
1999
5005403
769674
150821
1165
2697
2000
5254125
753257
158836
731
1493
2001
4886881
723503
171769
1002
1866
2002
5223283
727409
159463
872
1820
2003
5615238
777177
155857
810
1451
2004
6181251
850928
130278
841
1421
2005
6918605
944984
139464
1072
1374
2006
7818642
1036328
109563
1592
1953
2007
8505390
1116638
92087
1043
1550
2008
8802989
1170248
87960
865
1452
2009
9030202
1216308
86807
1171
1668
2010
10030810
1325749
98318
1444
1976
2011
11030939
1435879
91895
1700
2216
2012
11939620
1538006
77955
744
2213
2013
17129452
1611292
74847
1360
1694
2014
18097988
1679990
72367
1626
1509
2015
19053526
1740187
72344
1252
3596
2016
18367294
1749240
76668
1405
4642
2017
19511173
1874682
359653
1633
4226
2018
19374552
1879771
431276
1541
3773
Modelling with ANN consists of two stages as Training and Test. While trying to minimize the error
(deviation) value by controlling the input and output values given to the network during the training
process, it is desired to estimate the result by giving the input values without changing the weight values
in the test phase. The data set shown in Table 1 for the MLP network was devoted to training, 1970-
2009, for testing in 2010 and beyond.
Normalization of Data
The normalization process is used to eliminate the distances between data. In this study, the frequently
used normalization technique D_Min_Max normalization was preferred. Thanks to the normalization,
the data are normalized between 0.1 and 0.9. D_Min_Max normalization is as given in Equation 1.
(1)
Here, Xmax and Xmin show the maximum and minimum values of the data, respectively, and Xi shows
the current value of the data.
Network Performance Criterion
The performances of ANN models created in this study were compared using two different criteria.
These criteria are statistical parameters such as Mean Square Error (MSE) and Root Mean Square Error
TAM METİN SÖZEL SUNUMLAR
456
Square (RMSE). Mean Square Error (MSE) is one of the measurement methods used in digitizing the
difference between a regression curve's predicted value and actual value for a particular observation.
MSE value close to zero can be said to perform better. MSE value is calculated as in Equation 2.
(2)
Here, EJ represents the error that is the difference between the output produced by the network and the
actual output.
Root Mean Square Error Square (RMSE) is used to determine the error rate between measurement val-
ues and model estimates. The closer the RMSE value to zero is the higher ability to predict the model.
RMSE is the standard deviation of estimation errors (residues). The RMSE value is calculated as in
Equation 3.
(3)
Multilayer Perceptron (MLP)
Artificial Neural Network (ANN) has been recognized as a powerful tool to develop predictive models
for a complex context, such as human behaviour modelling and prediction, where relationships between
variables are highly complex and somewhat unknown. In their study, McCulloch and Pitts found that N
weighted parameters are entering the cells and the cells are going through a nonlinear function (McCul-
loch&Pitts, 1943: 115-133). Figure 1 is a simple neuro showing signal processing ability. It receives
signals from the input i (x1, x2,… xi), each with its connecting power (w1, w2,… wi). The neuron is then
multiplied by the weights of each input and the obtained sums are compared with the threshold value to
evaluate the signals. The calculated vector is then transformed using the activation function. Conse-
quently, the obtained value is the neuron's output (Ung et al., 2006: 339-356).
Figure 1. The artificial neuron [19]
The backpropagation algorithm is the method used to determine the weights in a multi-layer feed-for-
ward neural network to reduce the difference between target output and actual output. Interest in neural
networks begins with the development of the backpropagation algorithm by Rumelhart et al. (Rumelhart
et al., 1986: 318-352). Cybenko demonstrated that the power of the backpropagation algorithm can be
approached with a two-layer feed-forward network containing enough hidden neurons (Cybuko, 1989:
TAM METİN SÖZEL SUNUMLAR
457
303-314). During the training phase, the actual output is compared with the target output. If these two
outputs match, no changes are made to the network. If the requested match is not found, an error value
is generated. This error propagates back to the network and weights are adjusted accordingly.
The multi-layer perceptron (MLP) consists of a system of simple interconnected neurons or nodes, as
shown in Figure 2, a model representing a nonlinear mapping between an input vector and an output
vector (Gardner&Darling, 1938: 2627-2636). Multi-layer perceptron (MLP) are artificial neural network
systems with at least one hidden layer between the input and output layers. Since the data flows in these
sensors are forward, they fall into the feed-forward neural network class (Kargı, 2016).
Figure 2. Topological structure of multilayer perceptron network (He et al. 2019: 11)
IMPLEMENTATION
In this paper, a parametric study is carried out for the network structure that gives better results by
making experiments in MLP network structure about the data of 1970-2018 SSI. These structures are
tested with different parameters and an optimum network structure is obtained. The parameters we
changed for the MLP network include the learning rate, the number of neurons in the hidden layer, the
hidden layer, and the activation functions in the output layer. The learning rate ranges from 0 to 1 and
shows how quickly the model adapts to the problem. The hidden layer is the layer where forward calcu-
lations and backpropagation are made. The number of neurons in the layer may differ depending on the
problems. The activation function introduces the nonlinear real-world features of artificial neural net-
works. The learning rate has been tried by increasing 0.1 from 0.1 to 0.9. The number of neurons in the
hidden layer has been tried by increasing 10 from 10 to 100. As for the activation function, the most
used purelin, logsig, and tansig functions were used. The momentum factor (µ) was ignored during the
tests.
Network structure was created and tested on a system with Windows 10 operating system, 2 GB RAM
and i5 processor, and Matlab 2016b. As often used in the literature, the initial values of the weights are
assigned randomly. MLP network was tested by changing the parameters. Using these parameters, the
differences between the target output and the actual output are compared. There are different perfor-
mance measurement parameters in the literature. In this study, performance optimization was performed
according to MSE (Mean Square Error) and RMSE (Root Mean Square Error) values.
TAM METİN SÖZEL SUNUMLAR
458
Results of MLP Structure
A network structure with 5 inputs and 1 output has been created in the MLP network structure. For the
MLP network, the learning rate, the hidden layer, and the activation function in the output layer were
tested with 270 different results by changing the number of neurons in the hidden layer. There are 40
neurons in the hidden layer for the best output parameter values. A learning rate of 0.6 was used. A
logarithmic sigmoid function was used as an activation function in the hidden layer and output layer.
Following the given parameters, 0.2250 MSE and 0.1125 RMSE values were obtained. The network
structure was created with the MATLAB “newff” function, and the network structure is given in Figure
5.
Figure 5. MLP structure
Some of the parametric calculations performed for the MLP network and the number of neurons pro-
ducing the best output from the trials are presented in Table 2. Logarithmic sigmoid is used for the
activation function in the output layer. Since the output function cannot be negative for the data in this
study, the logsig function that produces positive output has been used (Tang et al., 2011: 185-794).
TAM METİN SÖZEL SUNUMLAR
459
Table 2. Some of the parametric calculations for MLP
# of Neurons in
Hidden Layer
Learning Rate
Activation
Function in Hid-
den Layer
MSE (test)
RMSE (test)
20
0.5
logsig
1.5291
0.7645
20
0.8
logsig
2.4850
1.2425
20
0.2
tansig
1.7173
0.8587
20
0.7
tansig
1.8227
0.9114
20
0.1
purelin
0.7190
0.3595
20
0.9
purelin
0.8078
0.4039
40
0.5
logsig
0.9334
0.4667
40
0.6
logsig
0.2250
0.1125
40
0.4
tansig
1.3860
0.6930
40
0.9
tansig
1.4788
0.7394
40
0.3
purelin
0.8671
0.4335
40
0.4
purelin
0.9271
0.4635
60
0.1
logsig
1.2064
0.6032
60
0.6
logsig
1.0748
0.5374
60
0.4
tansig
0.7673
0.3836
60
0.6
tansig
1.4937
0.7469
60
0.4
purelin
0.6807
0.3404
60
0.9
purelin
0.6340
0.3170
80
0.1
logsig
0.7785
0.3893
80
0.3
logsig
0.9516
0.4758
80
0.7
tansig
0.7839
0.3919
80
0.9
tansig
1.4174
0.7087
80
0.2
purelin
0.8586
0.4293
80
0.4
purelin
0.6760
0.3380
100
0.5
logsig
0.7927
0.3964
100
0.6
logsig
0.7081
0.3540
100
0.2
tansig
1.1968
0.5984
100
0.4
tansig
1.5405
0.7702
100
0.6
purelin
0.6676
0.3338
100
0.8
purelin
0.8166
0.4083
Figure 6 shows the graph between the output produced by the network and the actual output. This graph
shows the deviation between the actual output values and the estimation values for the disability esti-
mation according to the 0.2250 MSE and 0.1125 RMSE values we obtained. It can be created as a
graphic line with more output.
TAM METİN SÖZEL SUNUMLAR
460
Figure 6. Network output for test data actual output and estimate values
CONCLUSION
Occupational accidents are still a major problem for most companies today. In addition to health prob-
lems, work accidents cause extra costs and a decrease in the workforce, and delays in the project. Even
large-scale companies that have established occupational health and safety systems are likely to face
serious work accidents. Therefore, ANN was used to estimate the number of permanent incapacities of
the social security institution using compulsory insured numbers, number of workplaces, number of
occupational accidents, number of occupational casualties, and permanent incapacity to work data for
the years 1970-2018. While developing the ANN prediction model structure, MLP, the structure was
used. Experimental studies have been carried out to determine optimal training parameters such as the
number of neurons in the hidden layer and transfer function and learning rates. The network was then
trained and successfully tested with a 5-input 1-output data set from the data set. As a result of experi-
mental studies, it was seen that the MLP network structure gives quality results. In the continuation of
this study, studies on fuzzy neural networks or a little more expansion of parameters with more compre-
hensive data will be a part of future research.
REFERENCES
6331 Sayılı İş Sağlığı ve Güvenliği Kanunu
Abubakar, A. M., Karadal, H., Bayighomog, S. W., & Merdan, E., Workplace injuries, safety climate,
and behaviors: Application of an artificial neural network. International Journal of Occupational
Safety and Ergonomics, 1-11. https://doi.org/10.1080/10803548.2018.1454635, (2018).
Alaeddinoğlu, M. F., Sincar, S. & Naralan, A. İş Sağlığı ve Güvenliğinde Risk Analizi ve Değer-
lendirmesi için Geliştirilmiş Bir Karar Destek Sistemi (Yapay Sinir Ağı)-Atatürk Üniversitesi
TAM METİN SÖZEL SUNUMLAR
461
Örneği, Çukurova Üniversitesi Mühendislik Mimarlık Fakültesi Dergisi, 30(2), 275-291,
(2015).
Ayanoğlu C. & Kurt M., Metal Sektöründe Veri Madenciliği Yöntemleri ile Bir İş Kazası Tahmin
Modeli Önerisi, Ergonomics, 2(2), 78-87, (2019).
Ayhan B. & Tokdemir O., Keskin Bir Cisim ile Temas Sonucu Yaralanma Kazalarının Analitik
Hiyerarşi Prosesi ve Yapay Sinir Ağları ile Analizi, DÜMF Mühendislik Dergisi, 10(1), 323-
334, (2019).
Beriha, G. S., Patnaik, B., & Mahapatra, S. S. Assessment of occupational health practices in Indian
industries. Journal of Modelling in Management, 7(2), 180-200.
https://doi.org/10.1108/17465661211242804 , (2012).
Ceylan H. & Avan, M. Türkiyede’ki İş Kazalarının Yapay Sinir Ağları ile 2025 Yılına Kadar Tahmini,
International Journal of Engineering Research and Development, 4(1), 46-54, (2013).
Ceylan, H. An Artificial Neural Networks Approach to Estimate Occupational Accident: A National
Perspective for Turkey. Mathematical Problems in Engineering, 2014, 1-10.
https://doi.org/10.1155/2014/756326 , (2014).
Cybenko G., Approximation by superpositions of a sigmoidal function. Math. Control Signal Systems
2, 303–314, https://doi.org/10.1007/BF02551274, (1989).
Dizdar, E.N. & Koçar, O. İş Sağlığı ve Güvenliği Yönetim Sistemlerinde Risklerin Yapay Sinir
Ağlarıyla Değerlendirilmesi, Academic Platform Journal of Engineering and Science, 6-3, 73-
83, (2018).
Elmas Ç., Yapay Zeka Uygulamaları: Yapay Sinir Ağları - Makine Öğrenmesi - Derin Öğrenme - Derin
Ağlar - Bulanık Mantık - Sinirsel Bulanık Mantık - Genetik Algoritma, Seçkin Yayıncılık,
4.Baskı, Ankara, (2018).
Elmas, Ç. Yapay Sinir Ağları (Kuram, Mimari, Eğitim, Uygulama). 1. Baskı. Ankara: Seçkin Yayıncılık,
(2003).
Gardner, M. & Dorling, S. (1998). Artificial Neural Networks (The Multilayer Perceptron)-A Review
of Applications in the Atmospheric Sciences. Atmospheric Environment, 32(14-15), 2627–
2636. https://doi.org/10.1016/s1352-2310(97)00447-0
Goh, Y. M., & Chua, D. Neural network analysis of construction safety management systems: A case
study in Singapore. Construction Management and Economics, 31(5), 460-470.
https://doi.org/10.1080/01446193.2013.797095 , (2012).
He H., Yan Y., Chen T., & Cheng, P. Tree Height Estimation of Forest Plantation in Mountainous Ter-
rain from Bare-Earth Points Using a DoG-Coupled Radial Basis Function Neural Network, Re-
mote Sensing, 11(11), https://doi.org/10.3390/rs11111271, (2019).
Kargı V.S.A. Yapay Sinir Ağ Modelleri ve Bir Tekstil Firmasında Uygulama, Ekin Yayıncılık, Bursa,
ISBN: 978-605-5187-62-0. (2015).
TAM METİN SÖZEL SUNUMLAR
462
Karwowski, W., Zurada, J., Marras, W. S., & Gaddie, P. A Prototype of the Artificial Neural-Network
Based System for Classification of Industrial Jobs with Respect to Risk of Low Back Disorders.
In F. Aghazadeh (Author), Advances in industrial health and safety VI: Proceedings of the an-
nual International Industrial Ergonomics and Safety Conference held in San Antonio, Texas,
USA, 7-10 June 1994; the Official Conference of the International Foundation for Industrial
Ergonomics and Safety Research. London: Taylor & Francis, (1994).
Kho, S. M., Pahlavani, P., & Bigdeli, B. (2021). Classification and association rule mining of road col-
lisions for analyzing the fatal severity, a case study. Journal of Transport & Health, 23, 101278.
McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. The
Bulletin of Mathematical Biophysics, 5(4), 115–133. https://doi.org/10.1007/bf02478259
Moayed, F., & Shell, R. L. Application of Artificial Neural Network Models in Occupational Safety and
Health Utilizing Ordinal Variables. The Annals of Occupational Hygiene, 55(2), 132-142.
https://doi.org/10.1093/annhyg/meq079 , (2010).
Mohammadfam, I., Soltanzadeh, A., Moghimbeigi, A., & Savareh, B. A. (2015). Use of Artificial Neural
Networks (ANNs) for the Analysis and Modeling of Factors That Affect Occupational Injuries
in Large Construction Industries. Electronic Physician, 7(7), 1515-1522.
https://doi.org/10.19082/1515
Qiao, W., Liu, Y., Ma, X., & Liu, Y. (2020). A methodology to evaluate human factors contributed to
maritime accident by mapping fuzzy FT into ANN based on HFACS. Ocean Engineering, 197,
106892.
Recal, F., & Demirel, T. Comparison of machine learning methods in predicting binary and multi-class
occupational accident severity. Journal of Intelligent & Fuzzy Systems, (Preprint), 1-17.
Rumelhart, D. E., Hinton, G. E. &Williams, R. J. (1986). Learning internal representations by error
propagation, Parallel distributed processing: Explorations in the microstructure of cognition 1,
318–362.
Şen, Z. Yapay Sinir Ağları İlkeleri. İstanbul: Su Vakfı Yayınları, (2004).
Tang, L., Li, T., & Perng, C. S. LogSig: Generating system events from raw textual logs. International
Conference on Information and Knowledge Management, Proceedings, 785–794.
https://doi.org/10.1145/2063576.2063690, (2011).
Ung, S. T., Williams, V., Bonsall, S., & Wang, J. Test case-based risk predictions Werbos, P. J. (1988).
Generalization of Backpropagation with Application to a Recurrent Gas Market Model, Neural
Networks 1(4), 339-356. DOI: https://doi.org/10.1016/0893-6080(88)90007-X, (2006).
INTERNET RESOURCES
https://www.ilo.org/global/topics/safety-and-health-at-work/lang--en/index.html (E.T. 20.20.2020).
https://ilostat.ilo.org/data (E.T. 20.20.2020).
http://www.sgk.gov.tr/wps/portal/sgk/tr/kurumsal/istatistik/sgk_istatistik_yilliklari (E.T. 01.05.2020))
http://www.emo.org.tr/ekler/7dd0be2efb27309_ek.pdf (E.T. 22.03.2020)