Artificial Neural Networks and Advanced Fuzzy

Applied Artificial Intelligence An International Journal

ISSN: 0883-9514 (Print) 1087-6545 (Online) Journal homepage: http://www.tandfonline.com/loi/uaai20

Artificial Neural Networks and Advanced Fuzzy Techniques for Predicting Noise Level in the Industrial Embroidery Workrooms Mohsen Aliabadi, Rostam Golmohammadi, Hassan Khotanlou, Muharram Mansoorizadeh & Amir Salarpour To cite this article: Mohsen Aliabadi, Rostam Golmohammadi, Hassan Khotanlou, Muharram Mansoorizadeh & Amir Salarpour (2015) Artificial Neural Networks and Advanced Fuzzy Techniques for Predicting Noise Level in the Industrial Embroidery Workrooms, Applied Artificial Intelligence, 29:8, 766-785, DOI: 10.1080/08839514.2015.1071090 To link to this article: http://dx.doi.org/10.1080/08839514.2015.1071090

Published online: 24 Sep 2015.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at http://www.tandfonline.com/action/journalInformation?journalCode=uaai20 Download by: [mohsen aliabadi]

Date: 25 September 2015, At: 11:32

Applied Artificial Intelligence, 29:766–785, 2015 Copyright © 2015 Taylor & Francis Group, LLC ISSN: 0883-9514 print/1087-6545 online DOI: 10.1080/08839514.2015.1071090

Downloaded by [mohsen aliabadi] at 11:32 25 September 2015

ARTIFICIAL NEURAL NETWORKS AND ADVANCED FUZZY TECHNIQUES FOR PREDICTING NOISE LEVEL IN THE INDUSTRIAL EMBROIDERY WORKROOMS

Mohsen Aliabadi1 , Rostam Golmohammadi2, Hassan Khotanlou3, Muharram Mansoorizadeh3, and Amir Salarpour3 1 Department of Occupational Hygiene, Faculty of Public Health, Hamadan University of Medical Sciences, Hamadan, Iran 2 Department of Occupational Hygiene, Faculty of Public Health and Center for Health Researches, Hamadan University of Medical Sciences, Hamadan, Iran 3 Department of Computer Engineering, Faculty of Engineering, Bu-Ali Sina University, Hamadan, Iran

Noise prediction techniques are considered to be an important tool for evaluating cost-effective noise control measures in industrial workrooms. One of the most important issues in this regard is the development of accurate methods for analysis of the complex relationships among acoustic features affecting noise level in workrooms. In this study, artificial neural networks and advanced fuzzy techniques were employed to develop a relatively accurate model for noise prediction in the noisy process of industrial embroidery. The data were collected from 60 embroidery workrooms. Some acoustical descriptors of workrooms were selected as input features based on International Organization for Standardization (ISO) 11690-3. Prediction errors of all structures associated with neural networks and fuzzy models were approximately similar and lower than 1 dB. However, neurofuzzy models could slightly improve the accuracy of noise prediction compared with neural networks. These results confirmed that these techniques can be regarded as useful tools for occupational health professionals in order to design, implement, and evaluate various noise control measures in noisy workrooms.

INTRODUCTION Exposure to excessive noise, as the most important physical pollutant in industrial workrooms, can cause hearing loss, interfere with communication, disturb sleep, reduce performance, and provoke annoyance response and Address correspondence to Mohsen Aliabadi, Department of Occupational Hygiene, Faculty of Public Health, Hamadan University of Medical Sciences, P.O. Box 4171-65175, Hamadan, Iran. E-mail: [email protected], [email protected] Color versions of one or more of the figures in the article can be found online at www.tandfonline. com/uaai.


Neural Networks and Fuzzy Techniques for Predicting Noise in Workrooms

767

change in social behavior of workers (Attarchi et al. 2012). A hearing conservation program is considered to be the best approach to reduce the noise hazards in industrial workrooms to the lowest level (Kovacevic and Belojevic 2006). The basic elements of this program are prediction and determination of the noise level, which help find the most cost-effective noise control measures (Royster and Royster 1990). Prediction of noise induced by industrial processes, especially in the feasibility design of the process, can be most effective economically because planning costs should be kept as low as possible (Lewis 2002). It is noted that prediction of noise level in industrial processes helps evaluate the pre- and postprevention measures in terms of cost benefit and efficacy (Keränen and Hongisto 2010). In addition, industrial health professionals can use noise prediction tools to calculate noise exposure level in all parts of workrooms. It is, therefore, possible to compare this value with noise exposure limit in order to evaluate various solutions of noise control programs (ISO 11690-3 1997). A number of models have been developed for predicting the noise level in enclosed workrooms, ranging from simple empirical models to some more complicated and laborious models such as ray tracing (Elorza 2005; Heerema and Hodgson 1999). However, development of geometrical models is complicated and needs computer and acoustics expertise and time-consuming calculations, so they are frequently employed only in special cases. Acoustic features that influence noise propagation, due to the complexity of their nature, can harden the development of noise prediction models. The number of features that can influence the noise emission in fully enclosed spaces is numerous with nonlinear relationships (Elorza 2005). Comparison of simple room acoustic models applied to industrial workrooms showed that empirical models with high validity and applicability are rare (Hodgson 1996, 1998). The classic model of noise prediction in a closed workroom, called diffuse field theory, was proposed by Sabine based on the assumptions that the distribution of sound energy within a room is homogeneous, and that sound absorption is uniformly distributed. Mainly due to its simplicity, Sabine’s equation is normally used even though the real situations of workrooms seldom comply with its preconditions. The literature indicates that acousticians are not altogether satisfied with the existing equations, thus, this issue is still open to finding solutions better fitted to noise control applications (Hodgson 1998, 2003). One of the most important issues in this area is assigning more practical methods that can analyze the complex relationships among different features affecting noise level in enclosed workrooms. Presently, an artificial intelligence (AI) approach is assigned as an alternative statistical technique (Shank et al. 2008). This approach is recognized as a new method for accurately modeling various sciences. The most important


768

M. Aliabadi et al.

techniques of AI are artificial neural networks (ANNs), genetic algorithms (GAs), and fuzzy sets. During recent decades; fuzzy sets and NNs have offered interesting opportunities for analyzing the nonlinear and vague information located in the complex phenomena of the actual world (Konar 2000). ANNs can become knowledgeable via the process of acquiring empirical data. The ability to determine complex relations among various features and the capability to employ multiple learning algorithms are among their features (Basheer and Hajmeer 2000). Hence, ANNs can be a good choice for studying such physical phenomena as noise, for which adequate data related to various features are being collected although the mechanisms of interaction effects cannot be fully inferred (Paliwal and Kumar 2009). Note that, in the field of architectural acoustics, NNs have been used to predict the sound level of speech in classrooms; they have also been used to predict the noise level in opencast mines (Black et al. 2004; Nannariello, Hodgson, and Fricke 2001). Various problem-solving projects are excessively difficult to understand quantitatively. Fuzzy set theory, presented by Zadeh, resembles human reasoning in its use of approximate information and uncertainty to form decisions (Ross 2004). It was specifically developed to mathematically present uncertainty and vagueness and makes formalized implements for dealing with the imprecision intrinsic to many real situations (Fisher 2003). The results of applying fuzzy logic to the development of noise control models and acoustic comfort optimization in industrial workplaces suggest that fuzzy logic can be utilized to analyze very complex processes and can be successfully applied to highly nonlinear systems (Aluclu, Dalgic, and Toprak 2008; Zaheerudin and Jain 2008). Both NNs and fuzzy logic are principal techniques that have their strengths and defects. In order to have both learning and interpretation abilities in a unique system, integrating NNs and fuzzy systems to form neurofuzzy systems has been suggested (Yilmaz and Kaynar 2011). The combination of the two approaches into a united setup is regarded to be a favorable route toward developing an intelligent system that would be capable of mimicking natural characteristics of the human brain (Yilmaz and Kaynar 2011). Among the different neurofuzzy computing systems, adaptive neurofuzzy inference system (ANFIS) is a systematic approach for modeling and provides the best possible design specifications in a short time (RiahiMadvar et al. 2009). This advanced fuzzy inference technique has acquired great verification for modeling many real-world problems. In this regard, ANFIS can also be considered as a practical choice to facilitate development of the noise prediction model, which makes highly accurate analyses of acoustics conditions of workrooms possible. Considering the growing interest in developing noise prediction models with high level of validity and applicability, applying AI methods can be very


769

helpful and efficient. The noisy process of industrial embroidery, in which patterns and designs are sewn on cloth (Doll and Slomma 2002)is considered to be an important part of the textile industry. This study investigated the application of such AI approaches as ANN, generate fuzzy inference system (GENFIS), and ANFIS in developing a noise prediction model for the real-world situation of the industrial embroidery workrooms.


DESCRIPTION OF PREDICTION TECHNIQUE FEATURES In this study, the development of noise prediction techniques included different phases. The first phase consisted of selecting the appropriate research field in terms of noise pollution, recognizing likely features affecting the noise level of an industrial process, collecting and analyzing the dataset related to these features. The second phase consisted of determining types and parameters of various AI methods. The third phase included separating the data in order to train and test the dataset and developing noise prediction techniques. Finally, the validity of the best-developed models was assessed in terms of performance criteria using the test dataset. The data were collected from 60 industrial embroidery workrooms in the Khorasan province, east of Iran. Due to the nature of embroidery operations, the operators are exposed to excessive noise and are at some risk of hearing loss. The main acoustical features were determined based on ISO 11690-3 (ISO 11690-3 1997). The most important structural characteristics of workrooms were dimensions including length (L), width (W), height (H), total surface area (S), volume (V), and geometrical shape. Because the geometrical shapes of all workrooms were similar (all rectangular), this parameter was not considered as a candidate feature. The structural materials of workrooms were recorded carefully and their sound absorption coefficient values were specified from valid resources (Barron 2001). Because the surface absorption coefficient is a function of the frequency of the incidence sound, the noise reduction coefficient (NRC), which is defined as the average value of noise absorption coefficient in the 250 Hz, 500 Hz, 1000 Hz, and 2000 Hz octave bands, was used (Barron 2001). The total equivalent sound absorption area of workrooms in m2 units was calculated based on the ISO 12354-6 standard. The equivalent absorption area in a regular shaped room was determined as follows.

A=

n i=1

αS,i .Si +

o j=1

Aobj,j ,

(1)

770

M. Aliabadi et al.


where αs,i is absorption coefficient of surface i, Si is area of surface i, Aobj,j is equivalent absorption area of object j, n is the number of surfaces i; o is the number of objects j. Equivalent absorption surface area of objects located in workrooms mostly included embroidery machines surfaces and was also approximately estimated based on ISO 12354-6 recommendations (Neubauer and Kostek 2001). Average absorption coefficient of workroom (α (and room absorption constant (R) in m2 units and reverberation time (RT ) in seconds as the important acoustic features were determined based on Sabine’s theory according to Equations 2, 3, and 4, respectively (Barron 2001). A , St A , R= 1−α 0.16 × V , RT = A α¯ =

(2) (3) (4)

where St is the total surface area of each of workroom and V is the volume of each of workroom. The type of machinery maker (MT), number of embroideries (NE), number of embroidery heads (NH), lifetime of embroidery machine (EL), and operation speed of embroidery (ES) were features of embroidery machinery, which were recognized to be significant with regard to noise emission. Further, the most important feature of used raw materials, which could affect the noise level of embroidery operations, was the type of fabric (FT). The varieties of fabrics used were recorded in terms of decreasing thickness in seven major classes: three-layered fabric, felted, duplex cotton, lee, manoplex cotton, satin, and silk, based on the ordinal scale. Noise pressure level in the workrooms under study was considered as a target feature in the prediction techniques. Measuring equivalent noise level (Leq ) in workstations was performed based on ISO 9612 using integrated sound level meter Norsonic type132 (ISO 9612 2009). The Leq is the equivalent steady noise level of noise energy, averaged over time. Because noise is often a complex fluctuation, the noise level has to be averaged over a minimum sample time. The sampling time can be as short as a few minutes if the noise level is steady state. Therefore, the measurement time was considered to be 10 min. Regarding the length dimension of the machinery, three measurement points were considered as grid area in similar dimensions around the machinery. Then, logarithmic average of noise levels in reference points was calculated and recorded. Despite the uniformity of structural and acoustical features of each workroom, some new observations appeared in the workrooms by varying


771

TABLE 1 Descriptive Statistics of Final Features of Embroidery Workrooms for Developing Noise Prediction Techniques


Features Type Workroom Acoustic Features Average absorption coefficient Room absorption constant Reverberation time Embroidery Process Features Number of embroideries Number of embroidery heads Embroidery lifetime Embroidery speed Fabric typea Machine-maker typeb Target Feature Noise equivalent level a This b This

Minimum

Unit

α

−

0.04

0.14

0.06

0.02

R

m2

4.10

26.09

12.25

6.11

RT

s

0.94

4.21

2.70

0.89

NE NH

− −

1 6

4 40

1.24 14.31

0.53 5.70

EL ES

1 690

20 850

11.40 747

5.34 39.24

FT MT

Year Stitches/ Min − −

1 1

7 2

− −

− −

Leq

dB

85.20

1.96

79.40

Maximum

88.70

Mean

Std. Deviation

Symbol

feature was based on ordinal scale. feature was based on nominal scale.

features of the embroidery process and remeasuring the noise level. Hence, the number of possible recorded observations became 100. To process and select the final features of the noise prediction technique, statistical methods such as a correlation matrix were employed. Finally, nine features were recognized as final input features to develop the noise prediction technique. Descriptive statistics of the final structural, acoustical, and embroidery processes features in the workrooms are listed in Table 1. The distribution of noise level as the output feature in the studied embroidery workrooms is illustrated in Figure 1. The distribution of noise level in embroidery workrooms shows that the embroidery workrooms are noisy industrial processes. NEURAL NETWORKS STRUCTURE Multilayer perceptron, as used in this study, can be considered as the most widely accepted structure of NNs applied to physical phenomena. In terms of interconnections of structure of NNs, feed forward networks were considered. Training ANNs was performed based on the supervised learning method using an error backpropagation learning algorithm (Basheer and Hajmeer 2000). In this study, 80% of the dataset was randomly assigned to the training of the networks and 20% to the testing of the networks. Training the networks continued until the number of certain epoch


772

M. Aliabadi et al.

FIGURE 1 Histogram of noise-level distribution in embroidery workrooms.

(100) is passed or the error variations of network prediction in a mentioned epoch are minimized. Each input feature was normalized by subtracting it from its mean value and then dividing the result by its standard deviation in order to have zero mean value and unity variance for all variables. In this study, NNs have at least one hidden layer, with the number of neurons being approximately between half and twice the number of input features (4 and 17). The networks were developed with MATLAB software (Demuth and Beale 2002). GA is a search technique applied to determine true or rough solutions to optimization and search problems; it was employed to recognize the optimal values for initial weights of NNs (Fiszelew et al. 2007). In the next phase, principal component analysis (PCA) was employed as well to convert input features to new components without correlation in order to achieve a simple interpretation of information and improve the performance of the ANNs. ADVANCED FUZZY SYSTEMS STRUCTURE In fuzzy theory, fuzzy set A of universe X is described by function μA (x), named the membership function of set A. A = {(x, μA (x))|x ∈ X , μA (x) ∈ [0, 1]} ,

(5)



773

where μA (x) = 1 if x is totally in A; μA (x) = 0 if x is not in A; 0 < μA (x) < 1 if x is partly in A. For any element x of universe X , membership function μA (x) is the degree of membership to which x is an element of set A. A fuzzy inference system (FIS) consists of five components: (1) fuzzification of the input features, (2) implementation of a fuzzy operator such as AND, OR, and NOT in the antecedent, (3) an implication from the antecedent to the consequent, (4) aggregation of the consequents across the rules, and (5) defuzzification. A membership function (MF) is a curve that describes how each point in the input space is mapped to a degree of membership between 0 and 1. Fuzzy sets and fuzzy operators are the subjects and verbs of fuzzy logic. If–then rule statements are applied to formulate the conditional statements that compose fuzzy logic (The MathWorks 2001). Two advanced approaches of FIS employed for building noise prediction techniques are described. GENERATE FUZZY INFERENCE SYSTEM Generate fuzzy inference systems are the most important advanced fuzzy systems that can build from data using grid partition, subtractive clustering, and fuzzy c-means clustering. One of the most widely used advanced fuzzy systems for modeling complex phenomena with various features is GENFIS2. This method generates a Sugeno-type FIS structure using subtractive clustering and requires separate sets of input and output data as input arguments (The MathWorks 2001). Because there was not any definite knowledge of how many clusters could exist in the dataset, subtractive clustering, which is a fast algorithm for estimating the total number of clusters and the cluster centers in the dataset, was employed. After this step, the information from the clusters was applied to determine the initial number of rules and antecedent membership function used for constructing the FIS. The most important benefit of using the clustering method for rule generation is that the created rules are more suited to input data than they are in an FIS generated without clustering. The initial FIS was generated using the MATLAB fuzzy logic toolbox function GENFIS2 (The MathWorks 2001). In this study, a hybrid approach was employed that combined the advantages of the FIS and the GA in the unique system in order to achieve a highly accurate noise prediction model. Therefore, combinations of different AI approaches as complementary methods for predicting noise were investigated. As mentioned, the GA is a special type of evolutionary algorithm that uses techniques inspired by evolutionary biology, for example: inheritance, mutation, selection, and crossover or recombination (Szwarc and Czy˙zewski

774

M. Aliabadi et al.

2011). In this phase of the study, GA was also employed to select the best GENFIS2 in terms of the radius value of different parameters for each input feature. Fitness function of the GA is constructed by minimizing the value of error prediction based on a check dataset by GENFIS2. For this reason, 80% of the data were randomly selected and used to train networks and 20% were used to test the FIS equally.


ADAPTIVE NEUROFUZZY INFERENCE SYSTEM ANFIS is a prevailing tool in fuzzy modeling for memorizing knowledge about the given input and output datasets in order to calculate the membership functions that characterize the related FIS(Yilmaz and Kaynar 2011). Accordingly, in the final step of this study, ANFIS was applied to generate the best FIS system. ANFIS utilizes the result from GENFIS2 to begin optimization and uses a combined learning algorithm to enhance parameters of Sugeno-type fuzzy inference systems. It assigns a combination of the least squares method and the backpropagation gradient descent method for training FIS membership function parameters to copy a given training dataset. The first layer of ANFIS, as shown in Figure 2, is the input layer; the second layer characterizes the membership functions of each fuzzy input; the third layer is an inference layer, and normalizing is carried out in layer four. The fifth layer donates the output, and the sixth is the defuzzification layer. The layers include fixed and adaptive nodes; each adaptive node has a

FIGURE 2 Structure of typical adaptive neurofuzzy inference system.


775

resource of parameters and conducts a special function on entering signals (Boyacioglu and Avci 2010).


EVALUATION CRITERIA Root mean square error (RMSE) and coefficient of determination (R2 ) are the commonly used criteria for evaluating the performance of the prediction technique. These criteria determine the differences between predicted values and actual values of the issue. The RMSE and R2 criteria can be calculated as follows. n 2 j=1 (Tj − Yj ) , (6) RMSE = n (Tj − Yj )2 2 , (7) R =1− ( T j )2 2 Tj − n where Tj and Yj are the measured and predicted values of data j, respectively; and n represents the number of measurements. RESULTS Empirical techniques for prediction of noise in industrial embroidery workrooms were developed using AI approaches. The results showed that prediction errors of the various structures of ANNs were in the acceptable level, with RMSE lower than one dB for the unseen dataset. Based on these findings, networks structure with one hidden layer and 14 neurons exhibited the highest level of accuracy. Comparison of the different structures of the best ANNs in the training and testing phases based on RMSE and R2 are shown in Table 2. Furthermore, variations of prediction error of the network’s structure with one hidden layer in terms of the number of neurons in the training and the testing phases are presented in Figure 3. In different structures of the developed ANN model, using PCA for processing input features to new components could slightly decrease the prediction errors. TABLE 2 Comparison of Best ANN Prediction Models in the Training and Test phases Model Developing Phase

Train

Test

Evaluation Criteria

RMSE (dB)

R2

RMSE (dB)

R2

ANNs with PCA ANNs without PCA

0.67 0.79

0.90 0.84

0.69 0.84

0.88 0.79


776

M. Aliabadi et al.

FIGURE 3 Prediction errors of neural network structures in terms of the number of neurons for test data.

FIGURE 4 Comparison between measured and predicted Leq using ANNs with PCA for test data.

The plots of measured values of noise equivalent level versus predicted values by the best technique of NNs with one hidden layer and 14 neurons based on applying PCA in the training and the testing phases are illustrated in Figures 4 and 5. The performance of the developed advanced fuzzy inference system in the form of GENFIS and ANFIS in terms of RMSE and R2 are presented in Table 3. The results show that prediction error of a different structure of advanced fuzzy models was approximately similar. However, compared with GENFIS, ANFIS could slightly improve the accuracy of noise prediction. It is noted that accuracy of FIS techniques was also slightly higher than the ANN prediction technique. Due to the performance nature of the FIS model, prediction errors in the training phase of the developed FIS techniques were significantly low. The scatter plots of measured values of noise equivalent level versus predicted values by GENFIS and ANFIS are shown in Figures 6 and 7.



777

FIGURE 5 Comparison between measured and predicted Leq using ANNs without PCA for test data. TABLE 3 Comparison of Advanced Fuzzy Prediction Models in the Training and Test Phases Model Developing Phase Evaluation Criteria GENFIS ANFIS

Train

Test

RMSE (dB)

R2

RMSE (dB)

R2

0.00 0.017

1 0.99

0.55 0.53

0.87 0.88

FIGURE 6 Comparison between the measured and the predicted Leq using GENFIS for the test data.

The Gaussian membership function as shown in Figures 8 and 9 related to embroidery speed and number of embroideries was recognized to be the best membership function for the input features because of its simplicity and precision in determining the value of the input features. It is so adaptable


778

M. Aliabadi et al.

FIGURE 7 Comparison between the measured and the predicted Leq using ANFIS for the test data.

FIGURE 8 Membership functions related to embroidery speed.

to match with the nature of inputs. The fuzzy rule viewer of the established model, which is presented in Figure 10, indicates the behavior of the response (output feature) over the change in values of all input parameters. Figure 10 shows parts of 66 rules generated using this approach. Figures 11– 15 show the typical 3D illustration for different input features of the GENFIS prediction model along with the output features. The results show that the developed empirical techniques can predict the effectiveness of the different noise control interventions on the noise level of embroidery workrooms. Naturally, acoustics analysis of the typical embroidery workrooms based on the developed techniques can facilitate using the graphical user interface. Provision of a graphical user interface based on the best ANN, which is able to predict noise level of embroidery workrooms as a simple practical tool, was performed as shown in Figure 16.


779


FIGURE 9 Membership functions related to number of embroideries.

FIGURE 10 Typical rules graph of developed GENFIS for predicting noise level based on Sugeno approach.

DISCUSSION Highly accurate noise prediction techniques are considered to be an important tool for evaluating cost- effective noise control measures in industrial workrooms. Empirical techniques are more practical and quicker methods than other mentioned approaches for acoustic and occupational health experts. AI models are simple to apply and have generated a great deal of interest in engineering. By considering the successful application of this approach in complex engineering problems, in this study, AI methods were employed as an alternative approach in the field of acoustics modeling


780

M. Aliabadi et al.

FIGURE 11 Output in the form of 3D representation with embroidery machine speed and number of embroidery heads.

FIGURE 12 Output in the form of 3D representation with embroidery machine speed and embroidery machine-maker type.

to predict the noise level in closed workrooms, which can be regarded as a complex phenomenon. The analyses of results show that different prediction techniques could accurately predict the noise level in terms of acoustical parameters of workrooms and technical characteristics of real-noise sources (embroidery machines). The prediction errors of noise by the developed ANNs and GENFIS and ANFIS prediction models were lower than the subjective difference threshold for sound pressure level ±1 dB mentioned in ISO 3382 (ISO 3382-1 2008). The prediction error rate of developed techniques was at an acceptable level compared with similar studies that used AI to predict the sound level in architectural and environmental acoustics areas (Nanda, Tripathy, and Patra 2009; Black et al. 2004; Nannariello, Hodgson, and Fricke 2001).



781

FIGURE 13 Output in the form of 3D representation with embroidery lifetime and number of embroidery heads.

FIGURE 14 Output in the form of 3D representation with embroidery speed and number of embroideries.

The results confirm the high capabilities of AI approaches in improving the performance of acoustics prediction techniques compared with those of current empirical techniques developed using classical methods such as regression techniques in typical workrooms (Heerema and Hodgson 1999). Therefore, the developed empirical techniques can suitably predict the effectiveness of different scenarios (various solutions for noise control) on the noise level of embroidery workrooms. Although our technique achieved an acceptable level of validity in predicting noise level, the domain of application is restricted to studied fields. Noise levels in embroidery workrooms compared with national occupational exposure limit (OEL = 85 dB) demonstrates that most of embroidery workrooms are noisy industrial processes. Effective acoustic features such as average absorption coefficient of the


782

M. Aliabadi et al.

FIGURE 15 Output in the form of 3D representation with number of embroidery heads and number of embroideries.

FIGURE 16 Graphical user interface based on the best ANN.

workroom (α (for each workroom and embroidery-process features such as the number of embroidery heads (NH), lifetime of embroidery machine (EL), and operation speed of embroidery machine (ES) were recognized to be significant with regard to noise level. Due to the lack of an appropriate preventive maintenance system, the embroidery machines that were older produced the highest level of noise. For all textile machinery, some reductions in noise level can be achieved by preventive maintenance. Operation speed of embroidery heads can also influence the noise level of machines; the high speed of needles striking on the work surface can increase the noise level in workrooms. The type



783

of fabric used under the needles for embroidery operations also affects the noise level. In real situations, the fabrics with more strength and thickness can reduce the conduct area of the needles to the work surface so that quiet operations will be assured. In addition, subjective acoustic characters of embroidery workrooms were in range of live to medium-live due to using typical, hard constructional materials. By treatment of the ceiling and wall or even floor surfaces using sound-absorbing materials, it is possible to improve the acoustic performance of these workrooms and to reduce the share of the reflected noise level. The results confirmed the high capabilities of AI approaches in improving the performance of noise prediction techniques compared with those of current empirical techniques developed using classical methods such as multiple regression (Heerema and Hodgson 1999; Hodgson 1998; Norouzian and Asadpour 2012). It was found that one of the main causes for achieving this highly accurate prediction technique was the application of genetic algorithm as a complementary method that can determine the optimal structure of ANN and FIS models. The prediction error of networks showed a descending trend with an increase in the number of neurons, which can be due to using more nonlinear specifications and better learning of NNs. However, if the networks have too many hidden layers and neurons, they will follow the noise in the observations because of overfitting, resulting in weak generalization for untrained data (Obodeh and Ajuwa 2009). The input features of the ANN model, which could be likely to have higher correlation with each other, were analyzed using PCA. The results showed that the prediction error of the ANN model with PCA compared with the ANN model without PCA was slightly better. However, it has been noted that hidden layers of NNs can perform as feature extractors as analytical methods such as PCA do (Daayhoff 1996). The traditional method for constructing an FIS is based on human capabilities and expert knowledge. Therefore, this approach cannot be useful if the phenomenon is very complex, as is the case in noise pollution of an enclosed workroom. Presently, automatic approaches for constructing the FIS, such as GENFIS and ANFIS, are considered as advanced methods for removing these limitations in a short amount of time. ANFIS prediction technique had minimum prediction error compared with other developed methods. This result can be due to the best learning performance of ANFIS. In addition, comfort of implementation, speed and accuracy of learning, robustness of generalization capabilities, superior interpretation facilities through fuzzy rules, and ease of incorporation of both linguistic and numeric knowledge for problem solving are other advantages of the ANFIS method (Boyacioglu and Avci 2010). Finally, it was demonstrated that AI techniques can be used to construct accurate models for predicting noise in complex systems such as enclosed workrooms.

784

M. Aliabadi et al.

CONCLUSIONS


The accuracies of AI models for predicting the noise in industrial embroidery workrooms were within the acceptable level according to the international acoustics standard recommendations. AI approaches can, therefore, be regarded as good tools for minimizing the uncertainties in the field of industrial acoustic modeling, compared with the traditional methods. The developed prediction models can be regarded as useful tools for occupational health and acoustics professionals in order to design, implement, and evaluate various noise control measures in the noisy embroidery workrooms.

REFERENCES Aluclu, I., A. Dalgic, and Z. F. Toprak. 2008. A fuzzy logic-based model for noise control at industrial workplaces. Applied Ergonomics 39:368–78. doi:10.1016/j.apergo.2007.08.005. Attarchi, M., F. Dehghan, F. Safakhah, M. Nojomi, and S. Mohammadi. 2012. Effect of exposure to occupational noise and shift working on blood pressure in rubber manufacturing company workers. Industrial Health 50:205–13. doi:10.2486/indhealth.MS1321. Barron, R. F. 2001. Industrial noise control and acoustics. New York, NY: Marcel Dekker. Basheer, I. A., and M. Hajmeer. 2000. Artificial neural networks: Fundamentals, computing, design, and application. Journal of Microbiological Methods 43:3–31. doi:10.1016/S0167-7012(00)00201-3. Black, J., G. Benke, K. Smith, and L. Fritschi. 2004. Artificial neural networks and jobspecific modules to assess occupational exposure. Annals of Occupational Hygiene 48:595–96. doi:10.1093/annhyg/meh064. Boyacioglu, M. A., and D. Avci. 2010. An adaptive network-based fuzzy inference system (ANFIS) for the prediction of stock market return: The case of the Istanbul stock exchange. Expert Systems with Applications 37:7908–12. doi:10.1016/j.eswa.2010.04.045. Daayhoff, J. E. 1996. Neural networks architecture: An introduction. Boston, MA: International Thomson Computer Press. Demuth, H., and M. Beale. 2002. Neural network toolbox user’s guide for use with MATLAB. Natick, MA, USA: The MathWorks. Doll, V., and H. G. Slomma. 2002. Consideration of noise reduction and work place layout measures in the development of a new multiple-head embroidery machine. American Journal of Audiology 11:65–71. Elorza, D. O. 2005. Room acoustics modelling using the ray tracing method: Implementation and evaluation. Finland: University of Turku. Fisher, B. 2003. Fuzzy environmental decision-making: Applications to air pollution. Atmospheric Environment 37:1865–77. doi:10.1016/S1352-2310(03)00028-1. Fiszelew, A., P. Britos, A. Ochoa, H. Merlino, E. Fernandez, and R. Garcia-Martinez. 2007. Finding optimal neural network architecture using genetic algorithms. Researcher Computation Sciences 27:15–24. Heerema, N., and M. Hodgson. 1999. Empirical models for predicting noise levels, reverberation times and fitting densities in industrial workrooms. Applied Acoustics 57:51–60. doi:10.1016/S0003-682X(98)00037-1. Hodgson, M. 1996. When is diffuse-field theory applicable? Applied Acoustics 49:197–207. doi:10.1016/S0003-682X(96)00010-2. Hodgson, M. 1998. Experimental evaluation of simplified models for predicting noise levels in industrial workrooms. The Journal of the Acoustical Society of America 103:1933–39. doi:10.1121/1.421345. Hodgson, M. 2003. Ray-tracing evaluation of empirical models for predicting noise in industrial workshops. Applied Acoustics 64:1033–48. doi:10.1016/S0003-682X(03)00072-0.



785

ISO 11690-3. 1997. Acoustics – Recommended practice for the design of low-noise workplaces containing machinery – Part 3: Sound propagation and noise prediction in workrooms. Geneva, Switzerland: International Organization for Standardization. ISO 3382-1. 2008. Acoustics measurement of room acoustic parameters part 1: Performance spaces. Geneva, Switzerland: International Organization for Standardization. ISO 9612. 2009. Acoustics – Determination of occupational noise exposure engineering method. Geneva, Switzerland: International Organization for Standardization. Keränen, J., and V. Hongisto. 2010. Comparison of simple room acoustic models used for industrial spaces. Acta Acustica United with Acustica 96:179–94. doi:10.3813/AAA.918267. Konar, A. 2000. Artificial intelligence and soft computing behavioural and cognitive modelling of the human brain. Washington, DC: CRC Press. Kovacevic, M., and G. Belojevic. 2006. Tooth abrasion in workers exposed to noise in the montenegrin textile industry. Industrial Health 44:481–85. doi:10.2486/indhealth.44.481. Lewis, D. N. 2002. Application of indoor noise prediction in the real world. The Journal of the Acoustical Society of America 112:2267–2267. doi:10.1121/1.4779061. The MathWorks Inc.. 2001. Fuzzy logic toolbox for use with MATLAB user’s guide. Natick, MA: Author. Nanda, S. K., D. P. Tripathy, and S. K. Patra. 2009. Development of a neuro fuzzy model for noise prediction in opencast mines. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems 17:729–45. doi:10.1142/S0218488509006236. Nannariello, J., M. Hodgson, and F. Fricke. 2001. Neural network predictions of speech levels in university classrooms. Applied Acoustics 62:749–67. doi:10.1016/S0003-682X(00)00088-8. Neubauer, R., and B. Kostek. 2001. Prediction of the reverberation time in rectangular rooms with nonuniformly distributed sound absorption. Archives Acoustic 26:183–202. Norouzian, M. A., and S. Asadpour. 2012. Prediction of feed abrasive value by artificial neural networks and multiple linear regression. Neural Computing and Applications 21:905–09. doi:10.1007/s00521-011-0579-5. Obodeh, O., and C. I. Ajuwa. 2009. Evaluation of artificial neural network performance in predicting diesel engine NOx emissions. European Journal Sciences Researcher 33:642–53. Paliwal, M., and U. A. Kumar. 2009. Neural networks and statistical techniques: A review of applications. Expert Systems with Applications 36:2–17. doi:10.1016/j.eswa.2007.10.005. Riahi-Madvar, H., S. A. Ayyoubzadeh, E. Khadangi, and M. M. Ebadzadeh. 2009. An expert system for predicting longitudinal dispersion coefficient in natural streams by using ANFIS. Expert Systems with Applications 36:8589–96. doi:10.1016/j.eswa.2008.10.043. Ross, T. J. 2004. Fuzzy logic with engineering applications. West Sussex, United Kingdom: John Wiley & Sons. Royster, L. H., and J. D. Royster. 1990. Important elements and characteristics of hearing conservation programs and determination of their effectiveness. Environment International 16:339–52. doi:10.1016/0160-4120(90)90003-O. Shank, D. B., R. W. McClendon, J. Paz, and G. Hoogenboom. 2008. Ensemble artificial neural networks for prediction of dew point temperature. Applied Artificial Intelligence 22:523–42. doi:10.1080/08839510802226785. Szwarc, M., and A. Czy˙zewski. 2011. New approach to railway noise modeling employing genetic algorithms. Applied Acoustics 72:611–22. doi:10.1016/j.apacoust.2011.01.003. Yilmaz, I., and O. Kaynar. 2011. Multiple regression, ANN (RBF, MLP) and ANFIS models for prediction of swell potential of clayey soils. Expert Systems with Applications 38:5958–66. doi:10.1016/j.eswa.2010.11.027. Zaheerudin, and V. K. Jain. 2008. An expert system for predicting the effects of speech interference due to noise pollution on humans using fuzzy approach. Expert Systems with Applications 35:1978–88. doi:10.1016/j.eswa.2007.08.104.