predicting noise-induced hearing loss (nihl)

Predicting Noise-Induced Hearing Loss (NIHL) and Hearing Deterioration Index (HDI) in Malaysian Industrial Workers using GDAM Algorithm

Predicting Noise-Induced Hearing Loss (NIHL) and Hearing Deterioration Index (HDI) in Malaysian Industrial Workers using GDAM Algorithm M. Z. Rehman1, Nazri Mohd. Nawi2, M. I. Ghazali3 Faculty of Computer Science and Information Technology (FSKTM), Universiti Tun Hussein Onn Malaysia (UTHM), 86400 Batu Pahat, Johor. 1

Faculty of Mechanical and Manufacturing Engineering (FKMP), Universiti Tun Hussein Onn Malaysia (UTHM), 86400 Batu Pahat, Johor.

2,3

Email: [email protected], [email protected], 3 [email protected] ABSTRACT Noise is a form of a pollutant that is terrorizing the occupational health experts for many decades due to its adverse side-effects on the workers in the industry. Noise-Induced Hearing Loss (NIHL) handicap is one out of many health hazards caused due to excessive exposure to high frequency noise emitted from the machines. A number of studies have been carriedout to find the significant factors involved in causing NIHL in industrial workers using Artificial Neural Networks. Despite providing useful information on hearing loss, these studies have neglected some important factors. Therefore, the current study is using age, work-duration, and maximum and minimum noise exposure as the main factors involved in the hearing loss. Gradient Descent with Adaptive Momentum (GDAM) algorithm is proposed to predict the NIHL in workers. The results show 98.21% average accuracy between the actual and the predicted datasets and the MSE for both ears is 2.10x10-3. Hearing threshold shift found in the selected workers was greater than 25 dB, which means hearing impairment has occurred. Also, Hearing Deterioration Index (HDI) is found to be quite high for different sound pressure levels such as maximum exposure (dB) and average exposure (dB) but is reported normal for minimum exposure (dB) for all workers. KEYWORDS: hearing loss, hearing deterioration index, noise, occupational safety, noise-induced hearing loss.

1.0

INTRODUCTION

We hear different sounds in our daily life. And sometimes we are exposed to the sound without knowing the consequences of the exposure for a long period of time. When we hear noises that are too loud or it becomes painful to hear then a health condition called Noise-Induced Hearing Loss (NIHL) can occur (Stephen, 2002). NIHL is considered as the second most ISSN: 2180-3811 Vol. 3 June 2012 179 common form of sensori-neural hearing deficit after presbycusis (age

We hear different sounds in our daily life. And sometimes we are exposed Journal of Engineering and Technology to the sound without knowing the consequences of the exposure for a long period of time. When we hear noises that are too loud or it becomes painful to hear then a health condition called Noise-Induced Hearing Loss (NIHL) can occur (Stephen, 2002). NIHL is considered as the second most common form of sensori-neural hearing deficit after presbycusis (age related hearing loss) (Rabinowitz, 2000). NIHL is usually found greater in the developing industries of the world. NIHL is a common problem identified among the workers in the textile plants, basic metal industry, chemical industry, beverages and non-metallic mineral product industry. It was revealed in 1990’s Audiometric (hearing loss test) survey by Department of Safety and Occupational Health, Malaysia (DOSH) that about 26.9 percent of industrial workers had a hearing threshold of 30006000 Hz which was greater than normal and 21.9 percent of workers were already suffering from detectable hearing loss (Leong, 2003). A healthy Human ear makes it easier to hear and differentiate loud sounds from whispers. Any problem with the hearing ability damages the human’s life by reducing the quality of communication (Zaheeruddin & Jain, 2004). NIHL is a sensori-neural deficit that usually begins at the higher frequencies (such as 3000-6000 Hz) and develops gradually as a result of chronic exposure to excessive sound levels. NIHL can be stopped in earlier stages but in later stages hearing loss becomes permanent (Rabinowitz, 2000). Different studies have been done to detect NIHL in humans, but the recent improvements in Neural Networks have paved way for researchers to predict various harmful effects of noise on humans such as human work efficiency in noisy environment, noise induced sleep disturbance, speech interference in noisy environment, noise induced annoyance (Zaheeruddin, 2006) (Zaheeruddin, Jain, & Singh, 2006) (Zaheeruddin & Garima, 2005) (Zaheeruddin & Jain, 2004). In a recent study on NIHL by Yahya & Ghazali (2006), three variables such as age, work duration and noise exposure were selected and LevenbergMarquadt (LM) model was used for hearing impairment prediction in industrial workers. In another study, on tympanic membrane perforation, three factors were identified that directly affect human workers (i.e. noise 126 level, frequency and duration of exposure). It also negated the fact that age; an important factor in permanent hearing loss in older people can play the same effect on the young people (Zaheeruddin & Jain, 2004). Both studies on NIHL are in full-agreement that noise levels in excess of 90 decibels can cause permanent hearing loss to hair cells in the inner ear but still some important factors that can be helpful in finding harmful effects of NIHL in human hearing are neglected. The with2180-3811 NIHL detection is 2012 that mostly the input 180 main problem ISSN: Vol. 3 June parameters used by audiometric experts detecting NIHL are unclear and

decibels can cause permanent hearing loss to hair cells in the inner ear but still some important factors that can be helpful in finding harmful effects Predicting Noise-Induced Loss (NIHL) and Hearingare Deterioration Index (HDI) in Malaysian Industrial Workers using GDAM Algorithm of NIHLHearing in human hearing neglected. The main problem with NIHL detection is that mostly the input parameters used by audiometric experts detecting NIHL are unclear and not standardized as the data collected is often not precise and the environmental conditions are not suitable for the collection. For the sake of precision, this paper proposes a new algorithm to improve the working performance of Gradient Descent with Adaptive Gain (GDM-AG) model proposed by Nazri (2007) (Nazri, Ransing, & Ransing, 2007) that will change adaptively the momentum coefficient during the training. The proposed algorithm will be implemented using the input parameters (e.g. duration of exposure, age, minimum exposure, and maximum exposure) to predict NIHL and its effects on workers. The rest of the paper is organized as follows: the next sections describe the Artificial Neural Network (ANN), Back Propagation Neural Network (BPNN) and the effect of using the momentum coefficient in BPNN. Section-2.1.1, introduces the Gradient Descent with Adaptive Momentum (GDAM) algorithm for GDM-AG Model proposed by Nazri (2007). Results and Discussion of GDAM on NIHL data are discussed in Section-3 and finally the paper is concluded in the Section-4.

2.0 2.0

aRTIFICIAL NEURAL NETWORKS Artificial Neural Networks (ANNs) (ANNs)

Artificial Neural Networks (ANNs) are analytical techniques modeled on the learning processes of human cognitive system and the neurological functions of the brain. ANNs works by processing information like biological neurons in the brain and consists of small processing units known as Artificial Neurons, which can be trained to perform complex calculations (Deng, Chen, & Pei, 2008). 127 An Artificial Neuron can be trained to store, recognize, estimate and adapt to new patterns without having the prior information of the function it receives. This ability of learning and adaption has made ANN superior to the conventional methods used in the past. Due to its ability to solve complex time critical problems, it has been widely used in the engineering fields such as biological modeling, financial forecasting, weather forecasting, decision modeling, control systems, manufacturing, health and medicine, ocean and space exploration etc (Zheng, Meng, & Gong, 1992) (Kosko, 1994) (Basheer & Hajmeer, 2000) (Krasnopolsky & Chevallier, 2003) (Coppin, 2004) (Lee T. L., 2008).

An Artificial Neural Network (ANN) consists of an input layer, one or ISSN: Vol.of 3 neurons. June 2012 In ANN, every node 181 more hidden layers and an2180-3811 output layer in a layer is connected to every other node in the adjacent layer. ANN are

usually classified into several categories on the basis of supervised and unsupervised learning methods and feed-forward and feed-backward Journal of Engineering and Technology architectures (Deng, Chen, & Pei, 2008). Back-Propagation Neural Network (BPNN) is one of the most novel supervised-learning Artificial Neural Network (ANN) model proposed by Rumelhart, Hinton and Williams (Rumelhart, Hinton, & Williams, 1986). The BPNN learns by calculating the errors of the output layer to find the errors in the hidden layers. This qualitative ability makes it highly suitable to be applied on problems in which no relationship is found between the output and the inputs. Due to its high rate of plasticity and learning capabilities, it has been successfully implemented in wide range of applications (Lee, Booth, & Alam, 2005). Despite providing successful solutions BPNN has some limitations. Since, it uses gradient descent learning which requires careful selection of parameters such as network topology, initial weights and biases, learning rate, activation function, and value for the gain in the activation function. An improper use of these parameters can lead to slow network convergence or even network stagnancy. Previous researchers have suggested some modifications to improve the training time of the network. Some of the variations suggested are the use of learning rate and momentum to stop network stagnancy and to speed-up the network convergence to global minima. These two parameters are frequently used in the control of weight 128 adjustments along the steepest descent and for controlling oscillations (Zaweri, Seneviratne, & Althoefer, 2005). 2.1

BPNN with Momentum Coefficient

Momentum coefficient is a modification based on the observation that convergence might be improved if the oscillation in the trajectory is smoothed out, by adding a fraction of the previous weight change (Rumelhart, Hinton, & Williams, 1986) (Fkirin, Badwai, & Mohamed, 2009). So the addition of momentum-coefficient helps to smooth-out the descent path by preventing extreme changes in the gradient due to local anomalies (Sun, Zheng, Miao, & Li, 2007). In this case, it is essential to suppress any oscillations that results from the changes in the error surface (Norhamreeza, Nazri, & Ghazali, 2011). In earlier studies, static momentum-coefficient was found to be beneficial for the convergence to global minima but in later studies it was revealed that Back-propagation with Fixed Momentum (BPFM) shows acceleration results when the current downhill of the error function and the last change in weights are in similar directions, when the current gradient is in an opposing direction to the previous update, BPFM will cause the weight direction to be updated in the upward direction instead of down the slope as desired, so in that case it is necessary that the momentum-coefficient 182 ISSN: 2180-3811 June 2012 should be adjusted adaptively insteadVol. of 3keeping it static (Shao & Zheng, 2009) (Nazri, Rehman, & Ghazali, 2011).

that Back-propagation with Fixed Momentum (BPFM) shows acceleration results when the current downhill of the error function and the last change Predicting Noise-Induced Hearingare Loss in (NIHL) and Hearing Deterioration Index (HDI) the in Malaysian Industrial Workers using in weights similar directions, when current gradient is GDAM in anAlgorithm opposing direction to the previous update, BPFM will cause the weight direction to be updated in the upward direction instead of down the slope as desired, so in that case it is necessary that the momentum-coefficient should be adjusted adaptively instead of keeping it static (Shao & Zheng, 2009) (Nazri, Rehman, & Ghazali, 2011). In recent years several adaptive modifications of momentum are offered. In 1994, one such modification known as Simple Adaptive Momentum (SAM) (Swanston, Bishop, & Mitchell, 1994) was proposed to enhance the convergence capability of BPNN to global minima. SAM works by changing the momentum-coefficient according to the similarities between the changes in the weights at the current and previous iterations. If the weight change is in the similar ‘direction’ then the momentum is increased to speed-up convergence otherwise it is decreased. Although SAM was found as a better alternative to Conjugate Gradient Descent and conventional BPNN but it success and failure rate was to be same as conventional BPNN. C. Yu and B. Liu (2002) introduced a more efficient Back Propagation and Acceleration Learning (BPALM) method, to answer 129 the convergence failure problem in a much better way by adding some momentum to the adjustment expression. This can be accomplished by adding a fraction of the previous weight change to the current weight change. This encourages movement in the same direction on successive steps. The addition of momentum-step helps smooth-out the oscillations in the path by suppressing extreme changes in the gradient due to local anomalies. In 2008, R. J. Mitchell suggested adjusting the momentumcoefficient in SAM (Swanston, Bishop, & Mitchell, 1994) by considering all the weights in the Multi-layer Perceptrons (MLP). This technique of global adjustment of weights was found much better than the previously proposed SAM (Swanston, Bishop, & Mitchell, 1994) and helped improve the convergence rate to global minima (Mitchell R. J., 2008). In 2007, Nazri et al. proposed that by varying the gain parameter adaptively for each node can drastically progress the training time of the network. Based on Nazri et al. (2007) research, this paper proposes a further improvement on the algorithm that will use adaptive momentum and will keep the gain value fixed for all the trials. 2.1.1 Gradient Descent with Adaptive Momentum Algorithm(GDAM) In-order to increase the accuracy in the convergence rate and to make weight adjustments efficient on the current working algorithm proposed by Nazri et al., (2007) a new Gradient Descent Adaptive Momentum Algorithm (GDAM) is proposed in this section. ISSN: 2180-3811

Vol. 3

June 2012

183

In this paper, the modified Gradient Descent Adaptive Momentum

2.1.1 Gradient Descent with Adaptive Momentum Algorithm(GDAM) Journal of Engineering and Technology In-order to increase the accuracy in the convergence rate and to make weight adjustments efficient on the current working algorithm proposed by Nazri et al., (2007) a new Gradient Descent Adaptive Momentum Algorithm (GDAM) is proposed in this section.

In this paper, the modified Gradient Descent Adaptive Momentum (GDAM) algorithm is using batch mode of training for the entire training process. During the batch mode training all weights and biases and momentum are updated for the entire training set which is given to the network. The proposed algorithm, Gradient Descent Adaptive Momentum Algorithm (GDAM) adaptively changes the momentum while it keeps the gain value and learning rate fixed for the entire training. Mean Square Error (MSE) is calculated after each epoch and compared with the target error. The training continues until the target error is achieved or maximum For each epoch, epoch is reached. For each input vector, Step-1: Calculate the weights and biases using the previous momentum value. Step-2: Use the weights and biases to calculate new momentum value. For Endeach inputepoch, vector For each input vector, IF Gradient is increasing, increase momentum Step-1: ELSE decrease momentum Calculate End IF the weights and biases using the previous momentum value. Step-2: Repeat the above steps until the network reaches the desired value. Use the weights and biases to calculate new momentum value. End input vector IF Gradient is increasing, increase momentum ELSE decrease momentum End epoch End IF Repeat the above steps until the network reaches the desired value.

130

End epoch

∑(

(1) ( )

)

Adaptive ∑( Momentum is used to avoid oscillations in the network while searching the global minimum on the error surface. It smooth’s-out the ) path and helps the network in avoiding getting stuck in the local ( ) descent minima due to extreme changes in the gradient (Nazri, Rehman & Ghazali, Adaptive Momentum is usedgenerates to avoidaoscillations inweight the network while 2011) . Adaptive Momentum value for the updates in a searching the global minimum on the surface. It smooth’s-out the network. Here, the weight updates areerror limited to [0,1] as Log-sigmoid descent path and helps thetonetwork avoiding getting stuck in the local activation function is used find thein output on the jth node; minima due to extreme changes in the gradient (Nazri, Rehman & Ghazali, 2011) . Adaptive Momentum generates a value for the weight updates in a network. Here, the weight updates are limited to [0,1] as Log-sigmoid activation function is used to find the output on the jth node; 184

ISSN: 2180-3811

Vol. 3

June 2012

where,


where, where,

[∑ [∑

]

( (2) )

]

] [∑ ) Weights where, and biases are calculated in the same way, the weight (update expression ( ) ] links connecting to the output nodes with a bias is; [∑ for the Weights and biases are calculated in the same way, the weight update (3) (update ) expression forbiases the links a bias is; Weights and areconnecting calculatedtointhe theoutput same nodes way, with the weight ] [∑ ) ( ( expression for the links connecting to the output nodes with a bias is; ( ) Weights) and biases are calculated in the same way, the weight update ( ) ) ( ( expression for the links connecting to the output nodes with a bias is; ( ) Similarly, expression for the output nodes will be; ( ( ) bias) update Weights) and biases are calculated in the same way, the weight update ((4)) expression for) the connectingfor tothe theoutput outputnodes nodeswill withbe; a bias is; Similarly, bias update ( linksexpression ( ( ( ) bias) update ( ) Similarly, expression for the output nodes will be; ) ( ) ( ( Similarly, bias) update expression for the output nodes will be; (5) )  Where, is the momentum coefficient and is a positive number (in)the ) ( ( interval)[0,1]. For adjusting the momentum coefficient adaptively, gradient ( ) Where, is the coefficient and (g is snodes number in the Similarly, expression for gradient the output will be; path/trajectory is(momentum selected. When the ),a ispositive increasing (i.e. uphill), ) update (  bias interval [0,1]. For adjusting the momentum coefficient adaptively, gradient ) ( increase the else isdecrease momentum Where, is themomentum momentum coefficient coefficient and a positivethe number in )the path/trajectory is selected. the gradient (gs), isisincreasing uphill), coefficient. So,For here all theWhen momentum adjustment done with(i.e. respect to interval [0,1]. adjusting the momentum coefficient adaptively, gradient ) ( (  increase the momentum coefficient else decrease the momentum Where, momentum coefficient and is s),a ispositive number in the an increaseisorthe aisdecrease gradient. The momentum on each training at path/trajectory selected. in When the gradient (g increasing (i.e. uphill), ) So, here all the momentum adjustment is done with respect ( ) to coefficient. interval [0,1]. For adjusting the momentum coefficient adaptively, gradient epoch (s + 1) is calculated as follows: increase the momentum coefficient else decrease the momentum an increase So, or ahere gradient. The momentum on each training at path/trajectory isdecrease selected. When the gradient (gs), isisincreasing uphill), coefficient. all thein momentum adjustment done with(i.e. respect to Where,  is the momentum coefficient and is a positive number in the epoch (s + 1) is calculated as follows: increase the coefficientTheelse decreaseonthe an increase or amomentum decrease in gradient. momentum eachmomentum training at interval [0,1]. For adjusting the momentum coefficient adaptively, gradient coefficient. So, here all the momentum adjustment is done with respect epoch as follows: ( ) to { (s + 1) is calculated path/trajectory is selected. When the gradient (gs), is increasing (i.e. uphill), an increase or a decrease in gradient. The momentum on each training (6) at increase the momentum coefficient else decrease the momentum epoch (s + 1) is calculated as follows: ( ) to { coefficient. So, here all the momentum adjustment is done with respect where; ( ) at { an increase or a decrease in gradient. The momentum on each training epoch (s + 1) is calculated as follows: where; ( ) { Here in the equation (6), ‘p’ represents Here inthe thehighest equation and(6), the‘p’lowest represents valuethe in highest an where; the sigmoid interval, for example; the in sigmoid the interval interval, p = for [0.2,example; 0.9], 0.2 in is the the132interval p Here in and the equation (6),highest ‘p’ represents the highest and the lowestvalue value in to upda where; lowest 0.9 is the value lowest used and to 0.9 update is the momentum. highest After used ( ) { the sigmoid the interval, for example; in the p = [0.2, 0.9], 0.2update is the calculating Momentum Coefficient calculating in interval each the Momentum epoch, the weight Coefficient in132each epoc lowest and 0.9 is the highest value used to update momentum. After expression for the input node links expression becomes:for the input node links becomes: 132 where; the Momentum Coefficient in each epoch, the weight update calculating expression for the input node links becomes: 132 ( ) ( )] [∑ ( ) ( )] ( ( [∑ [∑

)(

)

(

)]

(

)

( )

132

( )nodes will be ) The bias update expression for hidden The bias nodes update willexpression be like this;for hidden ISSN: 2180-3811

Vol. 3

June 2012

The bias update expression for hidden nodes will be like this; ( ) ( ) ( ) (

185

)

( ) ) hidden ( )nodes ( )will ] expression (will be [∑ expression )( ) The ) like The bias update for bias update forthis; hidden nodes be like this; The bias update expression for hidden nodes will be like this; ( ) ) The bias update expression for hidden nodes will be like this; The bias update expression for hidden nodes will be like this; The bias update expression for hidden nodes will be like this; Journal of Engineering and Technology ( ) ) The bias update expression for hidden nodes will be like this; The bias update expression for hidden The bias nodes update willexpression be like this; for hidden nodes will be like The bias[∑ update expression ( ) for( hidden )]nodes ) (this; )] ( (will be like ( [∑ (( )) (( [∑ ))]] ( (( [∑ ( this;)] ( ( ) for( hidden)]nodes The bias[∑ update expression ( will )be like [∑ ( ) ( ) ( )] [∑ )]( ) ( ( ) ( )((] )) ( [∑ [∑ ()) (( )) ()) ) ( )] ) ( [∑ ( ) ) () )After ( calculating ( )layer ] ( biases, [∑ After calculating biases, hidden and )output layer)weight, hidden the andnet output weight weight, th After calculating biases, hidden and output layer weight, the net weight ( ) ) After calculating biases, hidden and output layer weight, the net weight After calculating biases, hidden and output layer weight, the n After calculating biases, hidden and output layer weight, the net weight becomes; becomes; becomes; ( ) ) After calculating biases, hidden and output layer weight, the net weight becomes; becomes; After calculating biases, hiddenAfter and output calculating layerbiases, weight, hidden the net and weight output layer weig becomes; Afterbecomes; calculating biases, hidden and output layer weight, the net weight becomes; becomes; becomes; After calculating biases, hidden and output layer weight, the net weight becomes; ( ) (( )) ( ) ) being u (with ) (after Finally, the network weights Finally, are the calculated networkafter weights beingare updated calculated Finally, the network weights are calculated after being updated (with ) Finally, the network weights are calculated after being updated with Finally, the network weights are calculated after being Finally, the coefficient, network weights are used calculated after being updated with momentum which momentum was coefficient, to dampen which the oscillations was used to the dampen in the upda osci momentum coefficient, which was used to dampen the oscillations the in Finally, the network weights are the calculated after being updated (with ) in momentum coefficient, which was used to dampen the oscillations the in with momentum coefficient, which was used to dampen the oscillati Finally, the coefficient, network weights are Finally, calculated network after being weights updated are calculated be momentum which was used to dampen the oscillations the the trajectory by adding previous the trajectory weight by change adding toprevious current weight weight change to after the curr the trajectory by adding previous change to the current weight in Finally, the network weights areweight calculated after being updated with momentum coefficient, which was used to dampen the oscillations the in the trajectory by adding previous weight change to the current weight in the trajectory by adding previous weight change to the current momentum coefficient, which was momentum used to dampen coefficient, the which oscillations was used the in to dampen the by adding previous weight change to current weight trajectory (Shao the trajectory & Zheng, the2009), trajectory the net (Shao weight & Zheng, becomes; 2009), the net weight becomes; the (Shao & 2009), the net weight becomes; momentum coefficient, which was used to& dampen the the in in Finally, network weights are calculated after being updated with the the trajectory by adding previous weight change tooscillations the current weight the trajectory trajectory (Shao & Zheng, Zheng, 2009), the net weight becomes; the trajectory (Shao Zheng, 2009), the netweight weight the by adding previous theweight trajectory change by adding to the current previous weight inbecomes; change to the trajectory (Shao & Zheng, 2009), the net weight becomes; the trajectory by adding previous weight change to the current weight in momentum coefficient, which was used to dampen oscillations the the trajectory (Shao & Zheng, 2009), the net weight becomes; the trajectory (Shao & Zheng, 2009), the trajectory the net weight (Shaobecomes; & Zheng, 2009), the net weight beco the trajectory & Zheng, 2009),weight the netchange weightto becomes; trajectory (Shao by adding previous the current weight ( ) in (( )) the trajectory (Shao & Zheng, 2009), the net weight becomes; ( ) ( )( ) ( ) 3.0 Results and Discussions 3.0 Results and Discussions 3.0 Results and Discussions 3. RESULTS AND DISCUSSIONS ( ) 3.0 3.0 Results and Discussions 3.0 Results Results and and Discussions Discussions The main focus of and thisDiscussions research The main is tofocus predict of this NIHL research in human is toindustrial predict NIHL in hum 3.0 Results The main focus this research to predict NIHL in human industrial 3.0 Results andof Discussions 3.0is Results Discussions The focus of this research to NIHL in human industrial The main focus ofmore thisand research is toHearing predict NIHL in human The main main focus ofDiscussions this research is to predict predict NIHL in and human industrial workers with more accuracy workers and is Hearing with Deterioration accuracy will be accurately Deterioration will 3.0 Results and workers with more accuracy and Hearing Deterioration will be accurately The main focus of this research is to predict NIHL in human industrial workers with more accuracy and Hearing Deterioration will be accurately workers with more accuracy and Hearing Deterioration will bein The main focus of this research The is to main predict focus NIHL of this in research human industrial is to predict NIHL workers with moreDiscussions accuracy and Hearing Deterioration be simulation accurately calculated among the selected calculated workers. among Before thediscussing selectedwill workers. the Before discussing t 3.0 Results and calculated among workers. Before discussing the simulation The main focus ofthe thisselected research is to Hearing predict NIHL in human industrial workers with more accuracy and Deterioration will be accurately calculated among the selected workers. Before discussing the simulation calculated among the selected workers. Before discussing the s workers with morethe accuracy and workers Hearing with Deterioration more accuracy willthe be and accurately Hearing Deterioration calculated among selected workers. Before discussing simulation test results there are certain things that need be explained such as tools workers with more accuracy and Hearing Deterioration will be accurately The results main focus ofare this research is to predict NIHL inselected human industrial calculated the selected workers. Before the test thereamong certain things that need bediscussing explained such as simulation tools calculated among the selected workers. calculated Before among thediscussing the workers. simulation Before discuss 133 and technologies, network topologies, testing methodology and dataset calculated among the selected workers. Before discussing the simulation workers with more accuracy and Hearing Deterioration will be accurately 133 and technologies, network topologies, testing methodology and dataset 133 133 used during the experimentation process. The discussion is as follows: calculated the selected workers. discussing thefollows: simulation used duringamong the experimentation process.Before The discussion is as 133 133 133

3.1 Preliminary Study 3.1 Preliminary Study 133 The Workstation used for the experimentation comes ready with a The Workstation used for the experimentation comes ready with a 2.33GHz Core-2 Duo processor, 1-GB of RAM and Microsoft XP (Service 2.33GHz Core-2 Duo processor, 1-GB of RAM and Microsoft XP (Service Pack 3) Operating System. Gradient Descent with Adaptive Momentum Pack 3) Operating System. Gradient Descent with Adaptive Momentum (GDAM) algorithm which is an improved version of the proposed (GDAM) algorithm which is an improved version of the proposed algorithm by Nazri (2007) is used to carry-out simulations on MATLAB algorithm by Nazri (2007) is used to carry-out simulations on MATLAB 7.10.0 software. 7.10.0 software. 3.1.1 NIHL dataset and Simulation Results 3.1.1 NIHL dataset and Simulation Results The Noise-Induced Hearing Loss (NIHL) dataset which consists of The Noise-Induced Hearing Loss (NIHL) dataset which consists of audiology study is obtained from Tenaga National Berhad (TNB), the audiology study is obtained from Tenaga National Berhad (TNB), the Electric Power Supply Company of Malaysia. The dataset consisting of Electric Power Supply Company of Malaysia. The dataset consisting of 1119 instances contains audiometric test information on each employee 1119 instances contains audiometric test information on each employee working from 1998 ISSN: to 2003 in TNB, 186 2180-3811 Vol.Terengganu 3 June 2012 State facilities. For working from 1998 to 2003 in TNB, Terengganu State facilities. For performing simulations, the dataset is divided into 900 subsets and 219 performing simulations, the dataset is divided into 900 subsets and 219

(GDAM) algorithm which is an improved version of the proposed algorithm by Nazri (2007) is used to carry-out simulations on MATLAB Predicting Noise-Induced Hearing Loss (NIHL) and Hearing Deterioration Index (HDI) in Malaysian Industrial Workers using GDAM Algorithm 7.10.0 software. 3.1.1 NIHL dataset and Simulation Results The Noise-Induced Hearing Loss (NIHL) dataset which consists of audiology study is obtained from Tenaga National Berhad (TNB), the Electric Power Supply Company of Malaysia. The dataset consisting of 1119 instances contains audiometric test information on each employee working from 1998 to 2003 in TNB, Terengganu State facilities. For performing simulations, the dataset is divided into 900 subsets and 219 subsets for training and testing respectively. Three layer back-propagation neural networks are used for testing of the models. The output is separated into two models i.e. left hearing loss and right hearing loss in order to reduce the complexity. Global Learning rate of 0.4 is selected for the entire tests and gain is kept fixed to 0.3. While log-sigmoid activation function is used as the transfer function from input layer to hidden layer and from hidden layer to the output layer. In this paper, the momentum term is varied adaptively between the range of [0, 1] randomly. For each problem, each trial is limited to 5000 epochs. A total of 30 trials are run for each momentum value to validate the best possible results. The network results are stored in the result file for each trial. Mean Square Error (MSE) and Correlation Coefficient(R) is used to verify the accuracy in the results. Table 1, shows the maximum accuracy, maximum CPU cycles, Mean Square error and best obtained value for Correlation Coefficient (R). The results demonstrated that the good performance in the training and testing sets indicates that the network is able to predict efficiently on the unseen data. 134 Table 1 Accuracy (%), Mean Square Error (MSE) and Correlation Coefficient (R) values for Hearing Loss Prediction in TNB Workers

NIHL Prediction for Both Ears Max CPU Cycles

902.5

Max Epochs

4367.5

Max Accuracy

98.21

MSE Correlation Coefficient (R)

2.10x 10-03 0.87

3.1.2 The Prediction Accuracy of GDAM To ensure the capability of the proposed Gradient Descent with Adaptive Momentum (GDAM)ISSN: Neural Network model, the prediction data and 2180-3811 Vol. 3 June 2012 187 actual data were compared to see the performance of the prediction. To

Correlation Coefficient (R)

Journal of Engineering and Technology

0.87

3.1.2 The Prediction Accuracy of GDAM To ensure the capability of the proposed Gradient Descent with Adaptive Momentum (GDAM) Neural Network model, the prediction data and actual data were compared to see the performance of the prediction. To evaluate the prediction performance of GDAM, the Correlation Coefficient (R) value is used. Four (4) workers belonging to middle-age groups are randomly selected here. The selected workers are Worker-1, Worker-2, Worker-3, and Worker-4. Referring to the Figure 1, the audiometric tests on the Worker 1 started at 38 years old and ended at age 42 years old. The actual data and prediction data comparison gave out the correlating values of 1.00 and 0.91 for Right and Left hearing. Figure 1 also illustrates GDAM prediction performance on worker 13. Worker 13 is working in TNB for the last 15 years and is showing a threshold shift of 25dB on both ears. According to the current noise standards, this person is considered already deaf. GDAM GDAMPerformance Performancefor for Actual ActualVs. Vs.Predicted PredictedData Datafor for Worker Worker1 1 Hearing Level (dB(A)) Hearing Level (dB(A))

2626 25.96 25.96 2525 2525 24.72 24.72 23.66 23.66 24.33 24.3323.33 23.33 2424 24.93 24.93 23.33 23.33 23.95 23.95 2323 23.32 23.32 24.86 24.86 21.74 2222 21.74 22.91 22.91 20.53 20.53 21.67 21.67 2121 19.89 19.89 2020 20.22 20.22 18.52 1919 18.52 19.34 19.34 1818 18.33 18.33 1717 1616 3838 3939 4040 4141 4242135 Right Right Hearing Hearing Actual Actual Level Level

18.33 18.33

19.34 19.34

20.22 20.22

23.33 23.33

2525

Right Right Hearing Hearing Prediction Prediction Level Level

18.52 18.52

19.89 19.89

20.53 20.53

22.91 22.91

24.86 24.86

Left Left Hearing Hearing Actual Actual Level Level

21.67 21.67

23.33 23.33

24.33 24.33

24.72 24.72

25.96 25.96

Left Left Hearing Hearing Prediction Prediction Level Level

21.74 21.74

23.66 23.66

23.95 23.95

23.32 23.32

24.93 24.93

2525

2525

2525

2525

2525

25dB 25dB Loss Loss

Figure Figure1 1 Noise-Induced Noise-InducedHearing HearingLoss Loss(NIHL) (NIHL)Prediction PredictionononRight Rightand andLeft LeftEars Ears for forWorker Worker1 1

Worker Worker2 2belongs belongstotothe themiddle-aged middle-agedgroup groupand andworking workingininTNB TNBfor forthe the last last1515years. years.During Duringthe theworking workinghours, hours,worker worker2 2isisexposed exposedtotomaximum maximum noise noiselevels levelsofof105.7 105.7decibels. decibels.Audiometric Audiometricresults resultsstates statesthat thatworker worker2 2 show showsimilar similarresults resultslike likeworker worker1 1and andhas hasdeveloped developeda ahearing hearingthreshold threshold shift shiftofofmore morethan than2525decibels decibelsasasshown shownininthe theFigure Figure2.2.The Thelines linesdepicted depicted ininFigure Figure2 2present presentthe thecritical criticalhearing hearingloss lossononboth bothears earsofofthe theworker worker2.2. 188 ISSN: 2180-3811 Vol. 3 June 2012 As Asexpected, expected,the thecompany companyalready alreadyhas hasa ahandicapped handicappedindividual individualininneed need

Worker 2 belongs to the middle-aged group and working in TNB for the last 15 years. During the working hours, worker 2 is exposed to maximum Predicting Noise-Induced Hearing Loss (NIHL) and Hearing Deterioration Index (HDI) in Malaysian Industrial Workers using GDAM Algorithm noise levels of 105.7 decibels. Audiometric results states that worker 2 show similar results like worker 1 and has developed a hearing threshold shift of more than 25 decibels as shown in the Figure 2. The lines depicted in Figure 2 present the critical hearing loss on both ears of the worker 2. As expected, the company already has a handicapped individual in need of immediate medical attention and guidance for future safety. Worker 3 is exposed to the 118 decibels of noise during the time period of 18 years. The audiometric tests information starts at age 43 and stops when the user is 47 years of age. The correlation coefficient (R) values for both ears are quite close to 1 which indicates a health trend between the actual and prediction data using GDAM algorithm. Although, there is one clear indication in the data which states that the hearing of both ears is continuously deteriorating with the passage of time. Figure 3, shows the hearing threshold shift of 33dB and 34 dB on both ears in Worker 3, which is quite higher than the 25 decibels threshold limit set by government authorities and calls for immediate attention to stop the further hearing deterioration. 136

GDAM Performance for Actual Vs. Predicted Data for Worker 2 26

Hearing Level (dB(A))

25 23 21

20.87

20

20

19

18.92

18

21.32 21 20.14 20

18.33

17

16

25.7 25.67 25.54

23.99 24.42 23.9 24 23.33 23.26 23.33

24 22

25.86

24.72

40

41

42

43

44

Right Hearing Actual Level

18.33

20

23.33

24

25.54

Right Hearing Prediction Level

18.92

20.14

23.9

24.42

25.86

20

21

23.33

24.72

25.67

20.87

21.32

23.99

23.26

25.7

25

25

25

25

25

Left Hearing Actual Level Left Hearing Prediction Level 25dB Loss

Figure 2 Noise-Induced Hearing Loss (NIHL) Prediction on Right and Left Ears for Worker 2

ISSN: 2180-3811

Vol. 3

June 2012

189


GDAM Actual Worker 33 GDAMPerformance Performancefor for ActualVs. Vs.Predicted PredictedData Datafor for Worker 3636

34.42 34.42 3434

Hearing Level (dB(A)) Hearing Level (dB(A))

3434

33.39 33.39 3333

3232 3030 26.62 26.62

2828 26.33 26.1826.33 2626 26.18

25.59 25.59

2424

2222 2020

21.33 21.33 21.29 21.29

24.15 24.15 22.85 22.85 22.56 22.56

2626 24.67 24.67

27.25 27.25

27.02 27.02

26.66 26.66 26.6 26.6

24.13 24.13

4343

4444

4545

4646

Right Hearing Actual Level Right Hearing Actual Level

21.33 21.33

22.56 22.56

24.67 24.67

26.66 26.66

3333

Right Hearing Prediction Level Right Hearing Prediction Level

21.29 21.29

22.85 22.85

24.13 24.13

26.6 26.6

33.39 33.39

Left Hearing Actual Level Left Hearing Actual Level

26.33 26.33

25.59 25.59

2626

27.02 27.02

3434

Left Hearing Prediction Level Left Hearing Prediction Level

26.18 26.18

24.15 24.15

26.62 26.62

27.25 27.25

34.42 34.42

2525

2525

2525

2525

2525

25dB Loss 25dB Loss

4747

Figure Figure3 3 Noise-Induced Noise-InducedHearing HearingLoss Loss(NIHL) (NIHL)Prediction PredictionononRight Rightand andLeft LeftEars Ears for forWorker Worker3 3

Figure Figure4 4represents representsthe thedata datarelated relatedtotoWorker Worker4 4who whoisisexposed exposedtoto110 110 decibels decibelsofofsound soundpressure pressureduring duringa atime timeperiod periodofof5 5years yearswhich whichstarts startsatat the age of 48 and end at the age of 52. The audiometric test results and the age of 48 and end at the age of 52. The audiometric test results andthe the prediction predictiondata datatrend trendare areshown shownininthe theFigure Figure4.4.Worker Worker4 4shows shows significant significanthearing hearingloss lossononboth bothears earsbut butstill stillthe theperson personisisworking workingininthe the same samenoisy noisyenvironment. environment.This Thisperson personcan canrecover recoverbybymeans meansofofartificial artificial hearing hearingaids. aids.The Thecorrelation correlationcoefficient coefficient(R) (R)values valuesininthis thiscase caseare are1.00 1.00for for Right Righthearing hearingand and0.90 0.90for forleft lefthearing hearingrespectively. respectively.

138 138

190

ISSN: 2180-3811

Vol. 3

June 2012


GDAM Performance for Actual Predicted for Worker GDAM Performance for Actual Vs. Vs. Predicted DataData for Worker 4 4 36

36 for GDAMPerformance Performance for ActualVs. Vs.Predicted PredictedData Datafor for Worker4 44 GDAM Performance for Actual Vs. Predicted Data for Worker GDAM Actual Worker

Hearing Level (dB(A))

Hearing Level (dB(A)) Hearing Level (dB(A)) Hearing Level (dB(A))

35 35 34 32.72 34 36 34.24 32.72 3636 32.67 32 32.67 3534.24 3535 3432 29.3630.1530.15 30.67 3434 30.67 32.72 32.72 29.36 32.72 30 30.67 30.67 34.24 34.24 32.67 34.24 32.67 32.67 3230 29 27.39 30.34 3232 30.15 27.45 29 30.15 27.39 30.34 28 29.36 30.15 30.67 29.36 30.67 27.45 30.67 29.36 30.67 30.67 30.67 27.33 3028 25.67 3030 27.33 27.51 25.67 26 27.39 29 27.39 30.34 27.39 30.34 2929 30.34 27.45 27.51 27.45 27.45 2826 2828 27.33 24.3 24.3 27.33 24 27.33 25.12 25.67 25.67 25.67 27.51 24 25.12 2624 27.51 27.51 20.67 2626 24 20.67 24.3 22 24.3 24.3 22 20.09 24 25.12 2424 25.12 20.09 20 20.67 24 25.12 20.67 2424 20.67 48 49 50 51 52 2220 2222 49 50 51 52 20.09 48 20.09 20.09 Right Hearing Actual Level 20.67 24 25.67 27.45 30.67 20 20 20 Right Hearing Actual Level 20.67 25.67 27.45 30.67 48 4924 50 51 52 49 50 51 52 4848 Right Hearing Prediction Level 20.09 49 24.3 50 25.12 51 27.51 52 30.34 Right Hearing Prediction Level 20.09 24.3 25.12 27.51 30.34 Right Hearing Actual Level 20.67 24 25.67 27.45 30.67 Right Hearing Actual Level 20.67 24 25.67 27.45 30.67 Right Hearing Actual Level 20.67 24 25.67 27.45 30.67 Left Hearing Actual Level 27.33 29 30.67 32.67 35 Left Hearing Hearing Prediction Actual Level 27.33 29 30.67 32.67 35 Right Level 20.09 20.09 24.3 25.12 27.51 30.34 Right HearingPrediction PredictionLevel Level 24.3 25.12 27.51 30.34 Right Hearing Left Hearing Prediction Level20.09 27.39 24.3 29.36 25.12 30.15 27.51 32.72 30.34 34.24 LeftHearing HearingActual Prediction 27.39 29.36 30.15 32.72 34.24 Left Level Level 27.33 27.33 29 30.67 32.67 35 Left HearingActual Actual Level 27.33 29 30.67 32.67 35 Left Hearing 25dB Loss Level 25 29 25 30.67 25 32.67 25 35 25 25dB Loss 25 25 25 25 25 Left Hearing Prediction Level 27.39 29.36 30.15 32.72 34.24 Left Hearing Prediction Level 27.39 29.36 30.15 32.72 34.24 Left Hearing Prediction Level 27.39 29.36 30.15 32.72 34.24 25dB Loss 25 25 25 25 25 25dB Loss 25dB Loss 2525 2525 2525 2525 2525

Figure 4 Noise-Induced Hearing Loss (NIHL) Prediction on Right Figure 4 Noise-Induced Hearing Loss (NIHL) Prediction on Right andand LeftLeft EarsEars for Worker 4 for(NIHL) Worker 4Prediction Figure4 44 Noise-Induced Noise-InducedHearing HearingLoss Loss (NIHL)Prediction Predictionon onRight Rightand andLeft LeftEars Ears Figure Noise-Induced Hearing Loss (NIHL) on Right and Left Ears Figure forWorker Worker4 44 for Worker for

3.1.3 Hearing Deterioration Index (HDI) 3.1.3 Hearing Deterioration Index (HDI) 3.1.3 Hearing Deterioration Index(HDI) (HDI) 3.1.3 Hearing Deterioration Index (HDI) There scientific correlations between noise levels, exposure, 3.1.3 Hearing Deterioration Index There areare scientific correlations between thethe noise levels, exposure, andand hearing damage risk. Extensive work undertaken in Dresden, Germany hearing damage risk. Extensive work undertaken in Dresden, Germany There arescientific scientific correlations between the noiselevels, levels, exposure, and There are scientific correlations between the noise levels, exposure, and There are correlations between the noise exposure, and (Kraak et al., 1977, 1981) shows the percentage risk of developing (Kraakdamage et al., risk. 1977, 1981) shows the percentage of developing hearing damage risk.Extensive Extensive work undertaken inrisk Dresden, Germanya a hearing damage risk. Extensive work undertaken Dresden, Germany hearing work undertaken inin Dresden, Germany hearing handicap and the median loss incurred with exposure as (Leong, hearing theshows median losspercentage incurred with as (Leong, et.handicap al. (Kraak et al.,1977, 1977,and 1981) showsthe the percentage riskexposure ofdeveloping developing (Kraak al., 1977, 1981) shows the percentage risk developing (Kraak etet al., 1981) risk ofof aaa 2003)(Bies & Hansen, 2003): 2003)(Bies & Hansen, 2003): hearing handicap and the median loss incurred with exposure as (Leong, hearinghandicap handicapand andthe themedian medianloss lossincurred incurredwith withexposure exposureasas(Leong, (Leong, hearing 2003)(Bies & Hansen, 2003): 2003)(Bies & Hansen, 2003): 2003)(Bies & Hansen, 2003): function of mean sound pressure level in the workplace (dBA) I. I.function of mean sound pressure level in the workplace (dBA) andand exposure (years). exposure (years). function ofmean meansound soundpressure pressurelevel levelinin inthe theworkplace workplace(dBA) (dBA)and and function mean sound pressure level the workplace (dBA) and I.I.I. function ofof II. function of hearing deterioration index, HDI. The formula is shown II. exposure function of hearing deterioration index, HDI. The formula is shown exposure (years). exposure (years). (years). below: below:ofof II. function function ofhearing hearingdeterioration deteriorationindex, index,HDI. HDI.The Theformula formulaisis isshown shown function hearing deterioration index, HDI. The formula shown II.II. below: below: below: [∫ [∫ [∫ [∫ [∫

l

] ]]

] dt]

where; where; where; where; where; is the mean exposure level (dBA), L isLthe mean exposure level (dBA), andand tthe ismean the time exposure. tisthe isthe time exposure. the exposure level(dBA), (dBA),and and mean exposure level (dBA), and LLLisis mean exposure level t is isthe thetime timeexposure. exposure. the time exposure. t tis

( () ) ( (( ) ))

139 139 139 139

ISSN: 2180-3811

Vol. 3

June 2012

191

139


It It is is evident evident that that the the hearing hearing deteriorates deteriorates very very rapidly rapidly during during the the first first 10 10 years and progressively more so as the exposure level rises above 80 dBA. years and progressively more so as the exposure level rises above 80 dBA. This implies implies that that to to avoid avoid hearing hearing impairment impairment in in 80 80 percent percent of of the the This population, the strategy that should be implemented is to avoid acquiring population, the strategy that should be implemented is to avoid acquiring aa HDI HDI greater greater than than 59 59 during during a a lifetime. lifetime. This This index index is is consistent consistent with with aa noise level level of of 85 85 dBA dBA exposure exposure over over a a lifetime. lifetime. At At 90 90 dBA, dBA, there there is is aa 20 20 noise percent risk of developing a hearing impairment after 30 years exposure percent risk of developing a hearing impairment after 30 years exposure as as shown in the figure 5. shown in the figure 5.

Figure Figure 5 5 Hearing Hearing Damage Damage as as aa function function of of Exposure Exposure (Bies (Bies & & Hansen, Hansen, 2003) 2003)

We We can can easily easily see see from from the the Table Table 2 2 that that all all of of our our workers workers who who have have been been working in working in the the factory factory for for 15 15 years years or or more more are are showing showing significant significant Hearing Hearing Deterioration Deterioration Indexes Indexes (HDI’s) (HDI’s) for for average average sound sound pressure pressure levels. levels. For average sound pressure levels, the workers are exceeding the HDI For average sound pressure levels, the workers are exceeding the HDI of of 59 and and same same is is the the case case with with the the Maximum Maximum sound sound Pressure 59 Pressure levels. levels. 140 140

192

ISSN: 2180-3811

Vol. 3

June 2012


Although, Although, it it shows shows less less HDI HDI but but still still the the hearing hearing losses losses are are quite quite significant here. All the workers are showing normal HDI for Minimum significant here. All the workers are showing for Minimum Although, it shows less HDI but still the normal hearingHDI losses are quite Although, it shows less HDI but stillcase theand hearing losses areduring quite sound pressure levels, but it is an ideal is only possible sound pressure it is anare ideal case and is only during significant here. levels, All thebut workers showing normal HDIpossible for Minimum significant here. All the workersboilers are showing normal HDI for Minimum the of timings and turbines. the start startpressure of operation operation timings of and turbines. sound levels, but it of is boilers an ideal case and is only possible during sound pressure levels, but it is an ideal case and is only possible during the start of operation timings of boilers and turbines. the start2of operation timings of boilers and turbines. Table Hearing deterioration Index for Workers Exposed to Different Sound Table 2 Hearing deterioration Index for Workers Exposed to Different Sound Pressure Levels Pressure Table 2 Hearing deterioration Index for Levels Workers Exposed to Different Sound Table 2 Hearing deterioration Index for Workers Exposed to Different Sound Pressure Levels HDI HDI HDI HDILevels HDI HDI Worker(s) Age T Worker(s) Age T Pressure (Avg. Exp.) (Min. Exp.) (Max. Exp.) (Avg. Exp.) (Min. Exp.) (Max. Exp.) HDI HDI HDI Worker(s) Age T HDI HDI HDI Worker 1 38 15 79.88 53.46 64.11 (Avg. Exp.) (Min. Exp.) (Max. Exp.) Worker(s) Age T Worker 1 38 15 79.88 53.46 64.11 (Avg. Exp.) (Min. Exp.) (Max. Exp.) Worker 38 15 79.88 53.46 64.11 Worker 1 2 40 15 80.73 54.41 64.41 2 40 80.73 54.41 64.41 Worker 1 38 15 79.88 53.46 64.11 Worker 2 40 15 80.73 54.41 64.41 Worker 43 18 85.9 56.20 71.95 Worker 3 32 40 15 80.73 54.41 64.41 43 18 85.9 56.20 71.95 Worker 3 43 18 85.9 56.20 71.95 Worker 48 19 83.68 56.14 67.88 43 18 85.9 56.20 71.95 Worker 34 4 48 19 83.68 56.14 67.88 Worker 4 Worker 4

48 48

19 19

83.68 83.68

56.14 56.14

67.88 67.88

A A comparison comparison of of HDI HDI in in workers, workers, when when they they are are exposed exposed to to different different levels of sound pressure in a working environment of TNB is shown in levels of sound pressure in workers, a workingwhen environment TNB is shown in the the A comparison of HDI in they areofexposed to different A comparison of HDI in workers, when they are exposed to different figure 6. figure of 6. sound pressure in a working environment of TNB is shown in the levels levels of sound pressure in a working environment of TNB is shown in the figure 6. figure 6. HDI HDI of of Workers Workers Exposed Exposed to to Noise Noise

Median Median Median Median Loss(dB) Loss(dB) Loss(dB) Loss(dB)

90.0 90.0 85.0 90.0 85.0 90.0 79.9 79.9 80.0 85.0 80.0 85.0 79.9 75.0 80.0 75.0 79.9 80.0 70.0 75.0 70.0 75.0 64.11 65.0 70.0 65.0 64.11 70.0 64.11 60.0 65.0 60.0 64.11 65.0 53.46 55.0 60.0 55.0 53.46 60.0 53.46 50.0 55.0 50.0 53.46 55.0 Worker 1 Worker 1 50.0 50.0 Worker 1 Worker 1

Figure Figure

HDI of Workers Exposed to Noise 85.9 HDI of Workers Exposed to Noise 85.9 80.7 80.7

85.9 85.9

80.7 80.7

83.7 83.7 83.7 83.7

71.95 71.95 71.95 71.95

64.41 64.41 64.41 64.41 54.41 54.41

56.20 56.20

67.89 67.89 67.89 67.89 56.14 56.14

56.20 56.14 54.41 56.20 56.14 54.41 Worker 2 2 Worker 3 3 Worker 4 4 Worker Worker Worker Percentage Risk of developing a hearing handicap Percentage handicap WorkerRisk 2 of developing a hearing Worker 3 Worker 4 HDI(Avg. Exposure)Worker 2 HDI (Min. Exposure) Worker 3HDI (Max. Exposure) Worker 4 HDI(Avg. Exposure) HDI (Min. Exposure) HDI (Max. Exposure) Percentage Risk of developing a hearing handicap Percentage RiskHDI of developing a hearing handicap HDI(Avg. Exposure) (Min. Exposure) HDI (Max. Exposure) 6 HDI of Workers exposed to different Pressure Levels HDI(Avg. Exposure) HDI (Min. Exposure) HDI (Max. Exposure) 6 HDI of Workers exposed to different Noise Noise Pressure Levels

Figure 6 HDI of Workers exposed to different Noise Pressure Levels Figure 6 HDI of Workers exposed to different Noise Pressure Levels 141 141 141 141

ISSN: 2180-3811

Vol. 3

June 2012

193


4.0

CONCLUSIONS Conclusions

Noise is a form of pollutant that is causing serious health problems to the workers inside the industry for many years. Continuous exposure to high pressure noise emitting from the machines can cause NIHL. In the developed countries noise is considered a serious threat to the health of blue collared employees. But in developing countries, noise is still not taken as seriously as it should be taken. In the recent years, many studies on NIHL are conducted in the developing countries. Following these studies, legislations are made but not taken seriously by the authorities in the industry and government agencies simply ignore to enforce these laws. The current research is carried out to detect NIHL in human industrial workers by using a proposed Gradient Descent with Adaptive Momentum (GDAM) algorithm to predict NIHL in workers using age, work-duration, and maximum noise exposure and minimum noise exposure. Overall, GDAM has shown outstanding results in-terms of predication accuracy on NIHL dataset collected from TNB. Mean Square Error (MSE) calculated for both ears is 2.10x10-3 while 98.21% average accuracy was achieved for the NIHL prediction. In the results it was found that Hearing threshold shift in all the workers was greater than 25 dB, which means hearing impairment has already occurred. Also, Hearing Deterioration Index (HDI) is found to be quite high for different sound pressure levels such as maximum exposure (dB) and average exposure (dB) levels but is reported normal for minimum exposure (dB) levels for all workers.

5.0

ACKNOWLEDGMENT

The Authors would like to thank Universiti Tun Hussein Onn Malaysia (UTHM) for supporting this research under the Fundamental Research Grant Scheme (FRGS), Vote No. 0737.

6.0

REFERENCES

B. Coppin. 2004. Artificial Intelligence Illuminated. Jones and Bartlet illuminated Series, USA. B. Kosko. 1994. Neural Network and Fuzzy Systems (1st Edition). Prentice Hall of India.

142

C. Yu, and B. D. Liu. 2002. A Backpropagation algorithm with adaptive learning rate and momentum coefficient. IJCNN, pp. 1218 - 1223. Honolulu.

D. Hansen. 2003. Engineering 194A. Bies, and H. C.ISSN: 2180-3811 Vol. 3 JuneNoise 2012 Control: Theory and Practice (3rd Edition). Spon Press, London, UK.

Predicting Noise-Induced Hearing LossNeural (NIHL) and Hearing Deterioration Index (HDI) in Malaysian Workers using GDAM B. Kosko. 1994. Network and Fuzzy Systems (1st Industrial Edition). Prentice HallAlgorithm

of India. C. Yu, and B. D. Liu. 2002. A Backpropagation algorithm with adaptive learning rate and momentum coefficient. IJCNN, pp. 1218 - 1223. Honolulu. D. A. Bies, and H. C. Hansen. 2003. Engineering Noise Control: Theory and Practice (3rd Edition). Spon Press, London, UK. D. E. Rumelhart, G. E. Hinton, and R. J. Williams. 1986. Learning Internal Representations by error propagation. Parallel Distributed Processing: Explorations in the Microstructure of Cognition. D. J. Swanston, J. M. Bishop, and R. J. Mitchell. 1994. Simple adaptive momentum: New algorithm for training multilayer Perceptrons. IEEE Journal of Electronic Letters, Volume 30 (18), pp. 1498-1500. Factories and Machinery Act 1967 (ACT 139) and Regulation and Rules (As at 30th 30th August 2011) (2011). Factories and Machinery (Noise Exposure) Regulations, 1989. PU (A) 106/89. F. Stephen. 2002. Continuous Noise-Induced Hearing Loss. Austin State University Press. Austin, USA. H. Shao, and H. Zheng. 2009. A New BP Algorithm with Adaptive Momentum for FNNs Training. GCIS-2009, pp. 16-20. Xiamen. H. Yan, Y. Jiang, J. Zheng, C. Ping, and Q. Li. 2006. A Multilayer PerceptronBased Medical Decision Support System for Heart Disease Diagnosis. Expert Systems with Applications, Volume 30 (2), pp. 272-281. H. Zheng, W. Meng, and B. Gong. 1992. Neural Network and its Application on Machine fault Diagnosis. ICSYSE 1992, pp. 576-579. I. A. Basheer, and M. Hajmeer. 2000. Artificial Neural Networks: fundamentals, computing, design and application. J. of Microbiological Methods,Volume 43 (1), pp. 03-31. K. Lee, D. Booth, and P. A. Alam. 2005. Comparison of Supervised and Unsupervised Neural Networks in Predicting Bankruptcy of Korean Firms. Expert Systems with Applications, Volume 29 (1), pp. 1-16. M. A. Fkirin, S. M. Badwai, and S. A. Mohamed. 2009. Change Detection Using Neural Network in Toshka Area. NSRC-2009, pp. 1-10. Cairo. M. N. Yahya, and M. I. Ghazali. 2005. Hearing Impairment Prediction on Malaysia Industrial Workers by using Neural Network. 8th International Conference on Quality in Research. Bali, Indonesia. M. S. Leong. 2003. Noise and Vibration Problems: How they affect us and the 143 Industry in the Malaysian Context. Skudai, Johor: University Teknologi Malaysia (UTM) Press, Malaysia. M. N. Nazri, M. R. Ransing, and R. S. Ransing. 2007. An improved Conjugate Gradient based learning algorithm for back propagation neural networks. ISSN: 2180-3811 Vol. 3 June 2012pp. 46-55. 195 Int. J. of Computational Intelligence, Volume 4 (1),

M. S. Leong. 2003. Noise and Vibration Problems: How they affect us and the Journal Industry of Engineering and Technology in the Malaysian Context. Skudai, Johor: University Teknologi Malaysia (UTM) Press, Malaysia. M. N. Nazri, M. R. Ransing, and R. S. Ransing. 2007. An improved Conjugate Gradient based learning algorithm for back propagation neural networks. Int. J. of Computational Intelligence, Volume 4 (1), pp. 46-55. M. N. Nazri. 2007. Computational Issues in Process Optimization using historical data. Ph.D Eng. Thesis, Swansea University, Swansea, United kingdom. M. N. Nazri, M. Z. Rehman, and M. I. Ghazali. 2011. Noise-Induced Hearing Loss Prediction in Malaysian Industrial Workers using Gradient Descent with Adaptive Momentum Algorithm. Journal: International Review on Computers and Software (IRECOS). Volume 6 (5). A. Norhamreeza, M. N. Nazri, and R. Ghazali. 2011. The effect of Adaptive Gain and Adaptive Momentum in improving training time of Gradient Descent Back Propagation Algorithm on Classification Problems. ICASEIT 2011, pp. 178-184. Bangi. P. M. Rabinowitz. 2000. Noise-Induced Hearing Loss. Journal of American Family Physician, Volume 61 (9), pp. 2759-2760. R. J. Mitchell. 2008. On Simple Adaptive Momentum. CIS-2008, pp. 01-06. T. L. Lee. 2008. Back-propagation neural network for the prediction of the shortterm storm surge in Taichung harbor, Taiwan. Engineering Applications of artificial Intelligence, Volume 21 (1), pp. 63-72. V. M. Krasnopolsky, and F. Chevallier. 2003. Some Neural Network applications in environmental sciences. Part II: advancing computational efficiency of environmental numerical models. Journal of Neural Networks, Volume 16 (3), pp. 335-348. W. J. Deng, W. C. Chen, and W. Pei. 2008. Back-propagation neural network based importance-performance for determining critical service attributes. Journal of Expert Systems and Applications, Volume 34 (2), pp. 1-26. Y. H. Zaweri, L. D. Seneviratne, and K. Althoefer. 2005. Stability Analysis of a Three-term Back-propagation Algorithm. Neural Networks, Volume 18 (10), pp. 1341-1347. Y. J. Sun, S. Zheng, C. X. Miao, and J. M. Li. 2007. Improved BP Neural Network 144 for Transformer Fault Diagnosis. Journal of China University of Mining and Technology, Volume 17 (1), 138-142. Zaheeruddin, G. V. Singh, and V. K. Jain. 2003. Fuzzy Modeling of Human Work Efficiency in Noisy Environment. Proceedings of the IEEE Int. conf. on fuzzy systems (FUZZ-2003), pp. 120-124. St. Louis, USA. Zaheeruddin, and V. K. Jain. 2004. An Intelligent System for Noise-induced Hearing Loss. Proceedings of International Conference on Intelligent Sensing and Information Processing (ICISIP – 2004), pp. 378-384. Chennai, 196 ISSN: 2180-3811 Vol. 3 June 2012 India.

Zaheeruddin, G. V. Singh, and V. K. Jain. 2003. Fuzzy Modeling of Human Work Predicting Noise-InducedEfficiency Hearing Loss (NIHL) and Hearing Deterioration IndexProceedings (HDI) in Malaysian WorkersInt. usingconf. GDAMon Algorithm in Noisy Environment. ofIndustrial the IEEE

fuzzy systems (FUZZ-2003), pp. 120-124. St. Louis, USA. Zaheeruddin, and V. K. Jain. 2004. An Intelligent System for Noise-induced Hearing Loss. Proceedings of International Conference on Intelligent Sensing and Information Processing (ICISIP – 2004), pp. 378-384. Chennai, India. Zaheeruddin, and Garima. 2005. Application of Artificial Neural Networks for the Prediction of Human Work Efficiency in Noisy Environment. CIMCA 2005, pp. 842-846. Vienna, Australia. Zaheeruddin, and V. K. Jain. 2006. A Fuzzy Expert System for noise-induced sleep disturbance. J. of Expert Systems with Applications , Volume 30 (4), pp. 761-771. Zaheeruddin, and V. K. Jain. 2005. Fuzzy Modeling of speech interference in noisy environment. ICISIP 2005, pp. 409-414. Zaheeruddin, V. K. Jain, and G. V. Singh. 2006. A Fuzzy Model for NoiseInduced Annoyance. IEEE Transaction on System, Man and Cybernatics, Volume 36 (4), pp. 697-705. Zaheeruddin 2006. Modelling of Noise-Induced Annoyance: A Neuro-fuzzy Approach. ICIT 2006, pp. 2686-2691. Mumbai, India.

145

ISSN: 2180-3811

Vol. 3

June 2012

197