Road Traffic Noise Prediction with Neural Networks-A Review

This paper aims to summarize the findings of research concerning the application of neural networks in traffic noise prediction. Modeling and prediction of traffic noise by means of classical approaches is a very complex and nonlinear process, due to involvement of several factors on which noise level depends. To overcome these problems, researchers and acoustical engineers have applied the artificial neural network in the field of traffic noise prediction. After a critical review of various neural network based models developed for road traffic noise prediction cited in the literature it was concluded that ANN based models were capable of predicting traffic noise more accurately and effectively as compared to deterministic and statistical models.


Introduction
Noise is one of the environmental pollutants encountered in daily life.Noise pollution has become a major concern of communities living in the vicinity of major highway corridors.In the view of rapid development, it is essential to study highway noise with respect to various causative factors.
With urbanization and the corresponding increase in the number of vehicles in metropolitan cities, pollution is increasing at an alarming rate.Main areas of concern are related to air and noise pollution.More than 70% of total noise in our environment is due to vehicular noise [1].Noise levels are showing an alarming rise and in fact the level exceeds the prescribed levels in most areas.Investigations in several countries in the past few decades have shown that exposure to noise, such as occurs when living in close proximity to busy roads or highways, has adverse effect on human health [2][3][4][5][6][7].
The level of highway traffic noise depends mainly on the following factors: i. Volume of the traffic ii.Speed of the traffic iii.Number of the heavy vehicles in the flow of traffic iv.Type of location v. Number of lanes present on highway vi.Road surface vii.Background noise etc.
To create a healthy and noise pollution free environment, a noise prediction model is needed so that the noise level along a busy highway can be predicted and investigated in advance during planning and design process [8].Traffic noise prediction models are required as aids in the design of highways and other roads and sometimes in the assessment of existing or envisaged changes in traffic noise conditions [9][10][11][12][13][14].These models are commonly needed to predict sound pressure levels, specified in terms of L eq. , L 10 etc., set by the government authorities.Neural networks offer a new strategy with enormous potential for many tasks in the planning, geographical and spatial.It is an area of interest how neural network addresses the shortcomings of conventional deterministic models.In the present manuscript an attempt has been made to explore the application of neural networks to road traffic noise prediction.

Artificial Neural Network
An ANN is an information processing paradigm that is inspired by the way a biological nervous system, such as the brain processes information.In this information processing system, the elements called neurons, process the information.It resembles the brain in two respects: i. Knowledge is acquired by the network through a learning process ii.Inter-neuron connection strengths known as synaptic weights are used to store the knowledge.An artificial neuron is characterized by: i. Architecture (connection between neurons) ii.Training or learning (determining weights on the connections) iii.Activation function All neural networks share some basic features.They are composed of simple processing elements, known as neurons.These elements take data from source as input and compute an output dependent in some well defined way on the values of inputs, using an internal transfer (i.e.activation) function.These neurons are joined together by some weights.Data flows along these connections and is scaled during transmission according to the values of weights as shown in Figure 1.The relationship between the input signals X 0, X 1... X n of neuron j and its output Y j is given by where W jo , W j1 ,…, W jn are the respective synaptic weights of neuron j, I j is the linear combiner output due to input signals and f is the activation function.
The arrangement of neurons into layers and the pattern of connection within and in-between layers are generally called the architecture of the net.The process of modifying weights according to the connections between the network layers, with the objective of achieving the expected output is called training a network.The internal process that takes place when a network is trained is called learning.Generally the three types of training are as follows: Supervised training: In a neural network, if for the training input vectors, the corresponding target output is known, the training method adopted is called supervised training.It is a process of providing the network with a series of sample inputs and comparing the output with the expected response.The training continues until the network is able to provide the expected response.
Unsupervised training: If for the input vectors, the corresponding target output is not known in a neural network, the training method adopted is called unsupervised training.A network can modify weights so that the most similar input vector is assigned to the output unit.The training process extracts the statistical properties of the training set and groups similar vectors into classes.
Reinforcement Learning: In this method, a teacher is also assumed to be present, but the right answer is not presented to the network.Instead, the network is only given an indication of whether the output answer is right or wrong.In this learning process, network attempts to learn the input-output mapping through trial and error with the view to maximize the performance index called the reinforcement learning signal.Reinforcement learning lies between supervised and unsupervised learning.It operates through continuing interactions between a learning system and the environment.
Weights: Weight is information used by the neural network to solve a problem.Neural networks consist of a large number of simple processing elements called neurons.These neurons are connected to each other by directed communication links, which are associated with weights.A bias acts exactly as a weight on a connection from a unit whose activation is always 1. where, Net = net input b = bias X i = input from neuron i W i = weight connecting neuron i to output.An activation function is used to calculate the output response of a neuron.The sum of a weighted input signal is applied to an activation function to obtain a response.Basically activation functions can be categorized in two types:-threshold and sigmoid functions.
Many types of ANN [15,16] have been developed such as the Kohonen network, the Learning vector quantization network, Grossberg's ART network, the Hopifield feedback network, the Competitive network and the Multi-layered feedforward network.Among these networks, the Multi-layered feed-forward network has been used extensively [17].A Multi-layered feed-forward network (MFFN) includes an input layer, a hidden layer and an output layer as presented in Figure 3.
MFFN are layered feed-forward networks typically trained with static back propagation.Their main advantages are that they are easy to use and that they can approximate any input/output map.

Review
Several attempts have been made to predict and model road traffic noise mathematically and statistically by different researchers.Due to the capability of neural networks to model nonlinear systems, Cammarata [18] studied the functional relationship between road traffic noise and related physical parameters.A back propagation network was applied to extract the functional relationship between particular road parameters (number of vehicles, average height of buildings, road width) and the level of sound pressure.The neural network approach was compared with those provided by a selected relationship found in the literature and showed that agreement between predictions and measurements are much more favorable to a neural network than to linear regression approach [19].
One limit of the neural method proposed, is its dependence on the acoustic measurements made.Errors in the measurement or incorrect interpretation make it impossible to supply (in the learning phase) the network with a set of data which can significantly represent the correlation between the parameters taken into consideration and the sound pressure level.In addition, during the production phase it is difficult to check whether a neural network output value (which is rather different from the value measured) is to be attributed to a defect in the learning phase or to an error in measurement.To present a strategy based on neural networks for the automatic recognition of acoustic measurements affected by errors they proposed [20] a neural approach to filter data provided by acoustic measurements.It was based on the use of a Kohonen Self-organizing Map network which receives correct acoustic measurements in the learning phase.The Kohonen neural network learning on the basis of this set of measurements would allow the network to be used as a filter.Having received a set of acoustic measurements in input, it would be able, in production phase, to discard any acoustic measurements which were insignificant or affected by errors.
Cammarata et al. [21] proposed use of a neural architecture made up of two cascading levels.At the first level a supervised classifying network, learning vector quantization (LVQ) network, filters data discarding all the wrong measurements, while at second level the BPN predicts the sound pressure level, which has been shown in Figure 4 and 5 respectively.Mark Dougherty [22] presented an interesting review based on application of neural networks to transport.Noise prediction models having acoustic energy descriptors, usually explicit as eq L or in two cases as a pseudo -10 L , were critically reviewed [23].Based on the traffic noise data measured near a highway in China, an artificial neural network was formed to predict further traffic noise levels.It was found that the predicted data by the ANN model were in good agreement with measured data and it was concluded that an artificial neural network provides a new method for traffic noise prediction [24].The position of the measurement station, the geographical situation between the noise source and the measurement station, wind speed and direction, air temperature and relative humidity, and time of day, [25] were considered in artificial neural networks to model the variation of noise levels from traffic around campus area.To have both the capabilities of learning and interpretability in networks and fuzzy logic are used for predicting the effects of noise pollution on human work efficiency [26,27].A neuro fuzzy computing system provides system identification and interpretability of fuzzy models and learning capability of neural networks in a single system.The neuro-fuzzy model was developed for predicting the effects of noise pollution on human work efficiency as a function of noise level, type of task, and exposure time.Noise generated from vehicular traffic is a major source of environmental pollution.A comprehensive study [28,29]  ) for the Tehran-Karaj express train.The model is developed upon the basis of the measurements 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.In the next step, the predictive capability of the model has been tested on the data associated with sampling locations, situated, respectively at distances of 35 m and 55 m from the centreline of track at a height of 1.5 m.It was observed that none of the models outweighs others as the best fit in a significant mode.After review of current literature on urban noise and Soft Computing techniques, performance of a neural network based system for the prediction of urban noise was described [31,32].An artificial neural network capable of predicting the level of urban noise, based on a set of 25 environmental characteristics was designed and implemented.Results reported by the network were compared with those of existing predictive models of urban noise, which were based on classical models.Tests confirmed that results produced by the network were better for all data records.Table 2 shows a comparison of classical models and the result obtained by neural network.Assuming bituminous road surfaces with mean traffic speeds of less than 75km/h, Givargis et al. [33] presented a basic and preliminary neural network model using a restricted database to predict Results indicate that a neural network approach can be applied for traffic noise prediction in Tehran in a statistically sound manner.The Wilcoxon matched-pairs signed-ranks test detects no significant difference between the absolute testing set errors of the developed neural network and a calibrated version of the CoRTN model.Table 3 shows absolute prediction errors given by their model.

Discussion
Two major advantages of ANN are that they are applicable to a wide variety of problems and relatively easy to use.In a neural network, relationships between variables are discovered automatically and fitting takes place naturally.Overall network structure is the only place where our intuition comes into play.Individual level data are necessary while using neural networks.The network operates on the data directly without the medium of an additional model.Like disaggregate models, neural networks also suffer from explanatory problems as there is a difficulty in interpreting the weights at this time.On the other hand, the fitting of parameters in a neural network is mathematically well founded.Neural networks provide a data tool through which we can model our intuition, without complications of having to formalize all complex causal variables and relationships with other models.
The level of highway traffic noise depends on many factors, such as the volume and speed of traffic, the number of heavy vehicles in the flow of traffic, condition of the vehicles and the road surface.All these parameters make noise modeling a complex process and nonlinear problem.ANN provides flexibility, massive parallelism, a learning and generalization ability, accuracy and some amount of fault tolerance in noisy and changing environments.
In the present review, it is concluded that Multilayered feed-forward network, which has been used extensively for noise reduction in ANN models, is the most mature one.It can also be concluded that back propagation is one of the most widely used paradigms surpassing (even) learning vector quantization and adoptive resonance theory in frequency of use.
We believe that detailed research and documentation of how neural based models in noise modeling and prediction are built will attract and encourage other persons working in this area.We also hope that this review will inspire a more methodological approach to emerge for modeling and prediction of traffic noise.

Conclusion
Problems related to noise prediction are highly complex and nonlinear in nature.Data sources are often numerous and complex.Neural networks are highly relevant to problems requiring large scale, highly dimensional, data analysis, such as noise prediction problems.Based on critical reviews (cited here), it was concluded that ANN has advantages over deterministic and statistical models, as it is well suited for implementation using very large scale integrated (VLSI) technology.But there still exist some problems, with ANN modeling such as its black box nature, convergence inaccuracy and highly data oriented nature.In order to overcome these problems, it is our thought that ANN should be implemented with other soft computing techniques such as fuzzy logic, particle swarm optimization and genetic algorithm.

Figure 3 .
Figure 3. Structure of feed-forward neural network

Figure 4 .Figure 5 .
Figure 4.The BPN Model given by Cammarata et al. [21] s roads at distances less than 4 m from the nearside carriageway edge and at a height of 1.4 m above the ground.They used the UK Calculation of Road Traffic Noise (CoRTN) approach.Overall model efficiency was examined using non-parametric tests, such as the Wilcoxon matched-pairs signedrank test for the training step and the Kolmogorove Smirnov test for two independent samples for the validation step.

Table 1 .
[28]ary of output from ANN Modeling for Interrupted traffic flow[28]