- Research article
- Open Access

# Modelling and evaluation of light railway system’s noise using neural predictors

- Selçuk Erkaya
^{1}Email author, - Abdurrahman Geymen
^{2}and - Bülent Bostancı
^{2}

**13**:20

https://doi.org/10.1186/s40201-015-0173-3

© Erkaya et al.; licensee BioMed Central. 2015

**Received:**26 September 2014**Accepted:**3 March 2015**Published:**17 March 2015

## Abstract

### Background

Noise is defined as a sound or series of sounds that are considered to be invasive, irritating, objectionable and disruptive to the quality of daily life. Noise is one of the environmental pollutants, and in cities it is usually originated from road traffic, railway traffic, airports, industry etc. The tram is generally considered as environmentally friendly, namely non-polluting and silent. However complaints from residents living along the tramway lines prove that it may sometimes cause annoyance. In this study, a Global Pointing System (GPS) receiver for determining the sampling locations and a frequency based noise measurement system for collecting the noise data are used to analyse the noise level in the city centre. Both environmental (background) and tram noises are measured.

### Results

Three types of neural networks are used to predict the noises of the tram and environment. The results of three approaches indicate that the proposed neural network with Radial Basis Function (RBF) has superior performance to predict the noises of the tram and environment.

### Conclusions

For making a decision about transportation planning, this network model can help urban planners for evaluating and/or isolating the tram noise in terms of human health.

## Keywords

- Noise mapping
- Neural networks
- Radial basis function

## Background

The first step of effective noise reduction is to define the main sources of the noise generation. Noise has a clear effect upon person’s health in terms of physiological and physical approaches. For example, negative effects on sense of hearing may be considered as a physical effect. These effects may cause temporary or permanent hearing loss. Physiological effects may include increase in blood pressure, cogwheel rigidity, stress, irregular heartbeat, pupil enlargement, insomnia, and tachyon etc. The application of neural networks has been rapidly expanded over the last two decades due to the progress in computer modelling and sensor technologies [1]. In case of system identifications using neural networks, the main purpose is usually to define a dynamically valid model which can be used for system analysis.

Modelling and prediction of noise by using classical approaches is a very complex and nonlinear process, due primarily to involvement of several factors on which noise level depends. To overcome these problems, researchers and acoustical engineers have applied the neural networks (NN) in the field of noise prediction. A literature review concerning the application of neural networks in traffic noise prediction was presented [2]. A back propagation neural network (BPNN) model was used to predict the noise caused by urban traffic. Five characteristics as the number of cars, trucks and motorcycles, the average height of buildings and the road width were used as input to model the equivalent sound pressure level as output [3]. Mathematical logarithmic, statistical linear regression and neural models were presented to predict the maximum A-weighed noise level (L_{A,max}) for the Tehran–Karaj express train. Measurements were obtained from sampling locations at distances of 25 m, 45 m and 65 m from the centreline of the track and at a height of 1.5 m [4]. Another study of same research group presented an artificial neural network model to predict hourly A-weighted equivalent sound pressure levels for roads in Tehran at distances less than 4 m from the nearside carriageway edge. Data were obtained from 50 sampling locations near five roads [5]. An experimental study was designed to assess the effect of road traffic noise on human performance [6]. A procedure for evaluating the noise quality of a cooling system (HVAC system), which frequently works during a day, was studied by using neural network structures [7]. Both experimental and simulation analyses were implemented together. For predicting the sound absorption coefficients of a sandwich structure nonwoven absorber, a general forecasting method was performed using general regression neural network (GRNN) [8]. Artificial neural network model was proposed to estimate the weighted sound insulation index value of wooden windows based on a limited number of windows parameters [9]. BPNN and GRNN models were used to predict the construction noise in Kuwait by considering the 33 construction sites. The GRNN model was superior to the BPNN model in its accuracy of predicting construction noise [10]. A neuro-adaptive active noise control approach with radial basis function neural network (RBFNN) was introduced for both modelling and controlling the system [11]. The feasibility and accuracy of network modelling for road traffic noise prediction was tested during the uninterrupted and interrupted traffic flow conditions [12]. Sixteen locations were identified by neural model comprising the feed forward negative back propagation algorithm. A procedure with neural predictor was proposed to test and evaluate the sound quality of cars with varying running speed [13]. Also, a robust noise analyser with NN was introduced to model and evaluate the joint noise of an industrial robot [14]. A Multi-Layered Perceptron (MLP) neural network model with Levenberg-Marquardt (LM) learning algorithm was developed for urban noise. 289 data from streets in Granada, Spain were obtained. The neural network model included 25 input variables to model the sound pressure level [15]. Noise measurements were made for 16 relevant outdoor points in the central campus area of Yildiz Technical University, Istanbul, Turkey. Artificial neural networks were used to model the variation of noise levels from traffic around the campus area. Six variables in input matrix, that is, sampling location, geographical situation, wind speed and direction, air temperature and relative humidity, and time of day were used to predict the sound pressure level [16].

The goal of this study is to present a neural network based approach to predict the level of environmental and tram line noises. Three types of neural predictor, which are extensively preferred in literature, are used for exact modelling. This paper is organized in the following manner. Section 2 outlines the methods, that is, working area, measurements and neural models. Results and discussion are given in section 3. Conclusions are outlined in section 4 for readers.

## Methods

### Working area and measurements

### Neural network

Three types of neural network were used to model and evaluate the noise characteristics of tram and environment. Structures of neural models are given in the following subsections.

#### Radial Basis Function Neural Network (RBFNN)

*x*

_{ i }) go from this layer to second layer. Three characteristics, that is, X, Y and Z coordinates of measured points, are considered in input matrix. In addition, tram type, tram velocity, air velocity and direction can be considered in input matrix for more stable network model. However, these characteristics have nearly zero changing. The sampling points are in the city centre and the traffic intensity is nearly similar in every hour for the daytime. Therefore, these effects were not considered in the current study. Hidden layer composed of RBF neural units is the second layer. The third layer is named as output layer, and the neurons of this layer have got linear transfer functions. Output matrix includes the tram and environmental noises. The neural network toolbox of MATLAB was used for developing the proposed network models [20]. Figure 3 outlines the RBFNN model. Hidden layer has got a non-linear Gaussian function,

*a*

_{ j }is defined as output of the

*j*

^{ th }node in the hidden layer.

*σ*

_{ j }is a width of the neuron

*j*,

*v*

_{ j }is given as Euclidean norm of the distance between input vector and neuron centre calculated as,

*c*

_{ j }is defined as a centre of the unit

*j*. The width of a unit is considered as the root mean square distance to the nearest unit

*j*. For the

*j*

^{ th }unit, the width \( {\sigma}_j \) is defined as

*c*

_{ 1 }

*, c*

_{ 2 }

*, . . . , c*

_{ ε }outline the nearest unit centres to the unit

*j*. The output value is given as

where *y*
_{
k
} defines the *k*
^{
th
} subsection of the *y* in the output layer, *d*
_{
jk
} is the weight between neuron of *j*
^{
th
} hidden layer and neuron of *k*
^{
th
} output layer. The number of hidden units is a crucial factor for determining the predictive properties of the network structure and is automatically calculated until the desired value of error is reached. In order to evaluate the best predictive property, various numbers of RBF units are used in network model. The number and width of RBF units are determined by hidden layer. After processing in hidden layer, the network model has got one weight value connected to the output layer. A linear squares regression algorithm is used to train the weight of output layer [21]. Mean square of the error (MSE) is considered as a performance measuring index of neural model.

#### Multi-Layered Perceptron Neural Network (MLPNN)

^{2}

*E*(

*ω*) is the Laplacian of the energy function and also referred to as the Hessian matrix. The Hessian term can be given,

*e*

_{ i }

*(ω)*is the error vector of the neural network for pattern

*i*and

*J(ω)*is the Jacobian matrix. By using the Taylor expansion,

*S(ω)*is nearly considered as zero for the GN method.

*S(ω)*comprises the second derivatives of the network error relative to the network weights. Since the number of computations increases exponentially with the size of the network, this term is very expensive for computing. The combination of the above equations gives the update rule of the GN method as,

If the *λ* is large, Eq. (9) approximates gradient descent. Otherwise, for a small *λ*, the algorithm approximates the GN method. If the *λ* is adaptively adjusted, the LM algorithm can make provision between its two extremes, that is, the gradient descent and the GN algorithm. Therefore, the advantages of gradient descent and the GN algorithms can be combined by LM method.

#### Generalized Regression Neural Network (GRNN)

*x*and

*y*. PDF can be derived from the data without any preconceptions about its form, the system is perfectly general. If

*f(x,y)*represents the known continuous PDF of a vector random variable,

*x*, and a scalar random variable,

*y*, the conditional mean of

*y*given

*X*is given by

*f(x,y)*is unknown, it must be predicted from a sample of

*x*and

*y*observations. The probability estimator \( \widehat{f}\left(X,Y\right) \) is based upon sample values

*X*

^{ i }and

*Y*

^{ i }of the random variables (

*x*and

*y)*;

*k*is the number of sample observations and

*p*is the x dimension

*.*A physical interpretation of the probability estimate \( \widehat{f}\left(X,Y\right) \) is that it assigns sample probability of width

*s*for each sample

*X*

^{ i }and

*Y*

^{ i }, and the probability estimate is the sum of those sample probabilities [22]. Defining the scalar function \( {D}_i^2, \)

Eq. (13) can be applied to problems comprising the numerical data.

## Results and discussion

According to the directive on the assessment and management of the environmental noise accepted in 2010 in Turkey, it is stipulated that the noise level (L_{day}) spread from the rail transport systems around during the day may not exceed the limit value of 65 dBA. However, by evaluating the Figures 4 and 5 together, 22 measurements at the 48 sampling points exceed this limit value.

As seen from Figure 10, convergence errors of the RBFNN are smaller than that of the GRNN and MLPNN predictors. This is a good reflection of modelling and predicting capabilities of the proposed RBFNN model. A good convergence was achieved by comparing the measurement results and the results from the RBFNN method. It can be concluded that the prediction of the environmental and tram noises is possible by using the RBFNN. Also, proposed neural model has a great accuracy to predict and estimate the noise characteristics under the received points. When the neglected parameters in input matrix, that is, tram type, measuring hour, tram velocity, air velocity or ground effects etc. have a considerable change, these characteristics should be considered in input matrix for more stable and accurate RBFNN predictor. These arrangements also affect the benefits and/or limits of neural model for noise prediction.

## Conclusion

In this study, a neural predictor is proposed to model and evaluate the characteristics of environment and tram noises. At the first stage of this study, the route of Kayseri light railway system, between 11.1 and 11.6 kilometres, is considered to evaluate the environmental and tram noise. 48 sampling points, where the traffic is highly intense in terms of the public transportation, are used for measuring. By considering the directive on the assessment and management of the environmental noise accepted in 2010 in Turkey, tram noises at some measured points (22 points) exceed the limit values (65 dBA).

At the second stage of this study, three types of neural networks, that is, multi-layered perceptron, generalized regression and radial basis function neural networks, are used for modelling and predicting the noise values. For training and testing of network structure, noise measurements are carried out to specify the network stability for exact modelling and predicting the noise characteristics. X, Y and Z locations for each point are taken into consideration in input matrix of network. In addition, the output matrix is constituted by environmental and tram noises. The simulation results give that the RBFNN has good convergence and high accuracy ability to predict the noise values at related points. This predictor is robust and stable to model the noise characteristics. On the contrary, other network models, that is, MLPNN and GRNN predictors are not suitable and should not be used to model and predict the noises correctly. The proposed neural predictor may be helpful to urban planners for making a decision about transportation planning. Also, these values can be used to select the correct materials for noise isolation.

## Declarations

### Acknowledgements

The authors wish to express their thanks to reviewer for their valuable contributions and Kayseri Transport Corporation-Kayseri Metropolitan Municipality.

## Authors’ Affiliations

## References

- Suzuki K. Artificial Neural Networks - Methodological Advances and Biomedical Applications. Rijeka: InTech; 2011.View ArticleGoogle Scholar
- Kumar K, Parida M, Katiyar VK. Road traffic noise prediction with neural networks-a review. Int J Opt Cont Theo Appl. 2012;2(1):29–37.Google Scholar
- Cammarata G, Cavalieri S, Fichera A. A neural network architecture for noise prediction. Neural Net. 1995;8(6):963–73.View ArticleGoogle Scholar
- Givargis SH, Karimi H. Mathematical, statistical and neural models capable of predicting LA, max for the Tehran-Karaj express train. App Acous. 2009;70:1015–20.View ArticleGoogle Scholar
- Givargis SH, Karimi H. A basic neural traffic noise prediction model for Tehran’s roads. J Env Manag. 2010;91:2529–34.View ArticleGoogle Scholar
- Alimohammadi I, Soltani R, Sandrock S, Azkhosh M, Gohari, MR. The effects of road traffic noise on mental performance. Iranian J Env Health Sci Eng. 2013; doi:10.1186/1735-2746-10-18.Google Scholar
- Erkaya S, Yildirim Ş. Evaluation of Noise Characteristics for a Cooling System Using Neural Network. 11th WSEAS Int Conf Robot Cont Manufac Tech and Multimedia Sys Signal Proces., 8–10 March 2011, Italy, 39–45.Google Scholar
- Liu J, Bao W, Shi L, Zuo B, Gao W. General regression neural network for prediction of sound absorption coefficients of sandwich structure nonwoven absorbers. App Acous. 2014;76:128–37.View ArticleGoogle Scholar
- Buratti C, Barelli L, Moretti E. Wooden windows: Sound insulation evaluation by means of artificial neural networks. App Acous. 2013;74:740–5.View ArticleGoogle Scholar
- Hamoda MF. Modeling of construction noise for environmental impact assessment. J Const Develop Count. 2008;13:79–89.Google Scholar
- Tokhi MO, Wood R. Active noise control using radial basis function networks. Cont Eng Prac. 1997;5(9):1311–22.View ArticleGoogle Scholar
- Parbat DK, Nagarnaik PB. Artificial Neural Network Modeling of Road Traffic Noise Descriptors. First Int Conf Emerg Trends in Eng Tech. 2008; doi:10.1109/ICETET.2008.220.Google Scholar
- Yildirim Ş, Erkaya S, Eski İ, Uzmay İ. Design of neural predictor for noise analysis of passenger car’s engines. J Sci Indust Resear. 2008;67(5):340–7.Google Scholar
- Yildirim Ş, Eski İ. Noise analysis of robot manipulator using neural networks. Robot Comp Integ Manufac. 2010;26:282–90.View ArticleGoogle Scholar
- Genaro N, Torija A, Ramos A, Requena I, Ruiz DP, Zamorano M. Modeling Environmental Noise Using Artificial Neural Networks. Ninth Int Conf Intel Sys Des Appl 2009; doi:10.1109/ISDA.2009.179.Google Scholar
- Avsar Y, Saral A, Gonullu MT, Arslankaya E, Kurt U. Neural network modelling of outdoor noise levels in a pilot area. Turkish J Eng Env Sci. 2004;28:149–55.Google Scholar
- Kayseri Transport Corporation, http://www.kayseriulasim.com. Accessed 25 Feb 2015.
- Brüel & Kjær Sound & Vibration Measurement, http://www.bksv.com. Accessed 25 Feb 2015.
- Broomhead D, Lowe D. Multivariable functional interpolation and adaptive networks. Complex Sys. 1988;2:321–55.Google Scholar
- MATLAB (ver 7.0), The MathWorks Inc., 3 Apple Hill Drive, Natick, MA 01760–2098.Google Scholar
- Erkaya S. Analysis of the vibration characteristics of an experimental mechanical system using neural networks. J Vib Cont. 2012;18:2059–72.View ArticleGoogle Scholar
- Specht DF. A general regression neural network. IEEE Trans Neural Netw. 1991;2(6):568–76.View ArticleGoogle Scholar
- Erkaya S. Prediction of vibration characteristics of a planar mechanism having imperfect joints using neural network. J Mech Sci Tech. 2012;26(5):1419–30.View ArticleGoogle Scholar

## Copyright

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.