Quantification of Acrylonitrile Butadiene Styrene Odor Intensity Based on a Novel Odor Assessment System With a Sensor Array

As an engineering plastic, acrylonitrile butadiene styrene (ABS) has been widely used in the automobile trims. The odor intensity of ABS can be considered as an important reference for the quality of in-vehicle air. Currently, many automobile manufacturers employ their own testing methods to measure the odor intensity of the trims. Different rules lead that the market lacks a unified standard to evaluate the odor intensity. In this paper, a novel odor evaluation system was proposed to measure the odor intensity of ABS. According to coefficient of variation (CV), analysis of variance (ANOVA), and principal component analysis (PCA), eight sensors were selected to compose an array with stability, repeatability, and selectivity. By means of the pretreatment and the feature extraction, the odor features were quantified by grey relation analysis (GRA). Then, the regression models were constructed by extreme learning machine (ELM), random forest (RF), and support vector machine (SVM) to predict the odor intensity. The results indicated that the quantified data could describe the odor intensity accurately and be predicted well by three models. This study demonstrated that the system achieved perception and quantification of the odor intensity of ABS. In conclusion, a self-developed system was put forward, offering a new technique to evaluate the odor intensity, prospective to replace the manual testing.


I. INTRODUCTION
With the development of automobile industry, the quality of in-vehicle air is getting more and more attention [1]. Despite people spend limited time in vehicles, the high concentration of pollutants in this microenvironment account for a large proportion of air pollutants [2]. For a nonsmoker, about 10-60 % of entire contaminants root in the exposure to vehicles [3]. The poisonous contaminants will endanger human health directly, leading to a series of diseases, such as allergic rhinitis, asthma, and even leukemia [4]. At present, pollution of in-vehicle air, hypertension, and AIDS have been listed as three threats to the human health by World Health Organization [5].
The associate editor coordinating the review of this manuscript and approving it for publication was Baozhen Yao .
The volatile contaminants release from the instrument panels, the door planks, the steering wheels, et al [6], [7]. Since favorable performances and lower cost, acrylonitrile butadiene styrene (ABS) is employed to process the automobile trims [8]- [10]. However, some monomers and emulsions may be remained during the course of polymerization, resulting in the diffusion of the irritant gases [10]. Besides, the trims are shaped from the master batches of ABS with high temperature and pressure. The materials release more volatile organic compounds (VOCs) when vehicles are exposed to the burning sun [6], [11]. At last, the pigments used for painting the trims emit the aromas of benzene series, and injure the respiratory tract and the skin under prolonged exposure to vehicles' microenvironment [11].
At present, the traditional physical and chemical analyses measure the odor intensity by detecting the concentrations VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/ of substances. However, there are some inevitable disadvantages. For instance, the colorimetry [12] is a suboptimal choice in long sampling period and slow response, and the electrochemical analysis [13] depends on strict operation environment. Despite the comprehensive two-dimensional gas chromatography [14] shows enhanced selectivity and improved sensitivity for the samples, the preparation of dirty samples is inevitable [15]. In conclusion, the traditional analysis methods ignore the measurement of comprehensive odor intensity. Consequently, auto manufacturers employ respective experts to test the odor intensity of automobile plastics. For example, Volkswagen heats and preserves the samples according to different temperatures and time. And the odor intensity is divided into six grades [16]. Nevertheless, General Motors utilizes the higher temperature to heat the samples and divide the odor intensity into ten grades [17]. Relative to above companies, TOYOTA detects the smells of parts from various regions and divides the odor intensity into six grades, from zero to five [18]. The detailed techniques and evaluation criteria are shown in Table 1. For different auto manufacturers, the pretreatment and grades of materials are inconsistent, but the manual testing is the main method to measure the odor intensity of ABS. Here, the test methods and evaluation criteria of three manufacturers are just listed. Despite artificial sensory evaluation is direct to sense the odor intensity [19], but causes damage to the health of the valuators working for a long time. Moreover, the valuators are probably affected by various factors, such as subjective preferences, working condition, and physical health [7], [20]. Above all, the automobile manufacturers and part plants employ their own valuators to measure the odor intensity of the same product, which gives rise to disputes for the quality of product. Simultaneously, different automobile manufacturers arise distinctions for the odor intensity, resulting in the lack of a unified standard in the whole market. The olfactory bionic technology applies gas sensor arrays [21], [22] to simulate the signals produced by olfactory cells, and combines with advanced recognition methods [23], [24]. The thin film on the sensor surface makes full contact and interaction (including physical adsorption, chemical adsorption, even chemical reaction) with the gases, and the electrical signals can be converted from chemical information [25]. Besides high accurateness and rapid responses, the sensors need to consider the cross sensitivity in the construction of the sensor array [26]. Szulczyński et al. [27] employed an electronic nose based on eight gas sensors to recognize the interaction of several components in the gas and surveyed the ingredients of the contaminants. On the other hand, the pattern recognition methods are employed to classify the signals representing the smell features. Osowski and Siwek [28] utilized the electronic nose to identify the distorted data of biological additives in the gasoline by means of principal component analysis (PCA), wavelet transform, support vector machine (SVM), etc., summing up the advantages of SVM in reducing errors. In the past few years, the olfactory bionic technology has progressed in the evaluation of the grades of paraffin [29], air quality [30], and the flavors of beverage [31]- [33].
This study proposed an odor evaluation system designed with a sampling device, a sensor array, and an analysis software to measure the odor intensity of four kinds of ABS. In order to construct the sensor array, coefficient of variation analysis (CV), analysis of variance (ANOVA), and PCA were applied to select the sensors with stability, repeatability, and selectivity. The features were extracted from the data obtained in the stable state. Then, grey relation analysis (GRA) [34] was implemented to quantify the odor intensity. Finally, the regression models were established by means of extreme learning machine (ELM), random forest (RF), and SVM to predict the quantified values of unknown samples. And the quantified values have a good performance in the training sets (R 2 > 0.91) and test sets (R 2 > 0.88). In conclusion, the odor intensity of ABS detected by the odor evaluation system was described in terms of the accurate numbers, rather than rough grades. Therefore, the self-developed evaluation system provided a novel technique to measure the odor intensity of ABS, promising to unify the odor evaluation criteria.

A. SAMPLES PREPARATION
In this study, four kinds of ABS with the same grade were purchased from Qi Mei Company (a manufacturer of ABS), which were high heat-resistant ABS, heat-resistant ABS, flame-retardant ABS, and polypropylene and glass ABS. Different ABS brands were stored in distinct storage bags to avoid pollution from other smells. Additionally, the disposable gloves were required to be worn before sampling.
To begin with, weigh the particles of ABS according to 10g, 20g, 30g, 40g, and 50g. Then, the samples were placed in bottles with 200 ml and sealed with lids. Eighty tests were implemented for each kind of ABS (e.g. Brand high heat-resistant ABS: 10g × 16, 20g × 16, 30g × 16, 40g × 16, 50g × 16), and a sum of 320 samples could be obtained. After sampling, the bottles were preserved in a cool and dry environment.

B. DESIGN OF ABS ODOR EVALUATION SYSTEM 1) ABS ODOR EVALUATION SYSTEM
The odor evaluation system consisted of a sampling device, a sensor array, and an analysis software. The whole construction of the system is shown in Fig. 1. And Fig. 2 shows the photograph of the whole system and the interior detection device.
The sampling device was designed in accordance with dynamic headspace sampling, including an air generator and a thermostatic heater. The air generator provided pressure for the circulation of the gases, where the active carbon was used to wipe off the vapors and the impurities from air. The thermostatic heater was employed to generate the volatile gases by heating the particles of ABS at 40 • C.
As for the gas chamber, made up of poly tetra fluoroethylene (PTFE), a synthetic polymer material that is insoluble in almost all organic solvents. The structure of the gas chamber is shown in Fig. 2 (b). The sensor array was integrated by ten sensors. In this paper, eight sensors were applied to detect the odors, which were MQ-2, MQ-3, MQ-137, TGS-813, TGS-821, TGS-825, TGS-830, and TGS-2610, selected from Table 2. The valves and pumps were applied to control the cleaning of the gas circuit and gas chamber. The EM936M (a data acquisition card) collected the multi-channel sensor response signals and converted the analog signals into digital signals. At last, the signals were transmitted to a PC.
The analysis software was compiled on C programming language to accomplish the start-up of the equipment, and the data communication and storage. At the same time, the implanted MATLAB software was implemented to complete the pre-processing and the feature extraction of raw data, and quantify the odor intensity.

2) SENSORS SELECTION
In the construction of the sensor array, the sensors were supposed to be selected with stability, repeatability, and selectivity when responding to the gases. On the other hand, the high-dimensional data may result in the redundant features. Thus, based on four finds of ABS, the steps of selecting the sensors are as follows: a: COEFFICIENT OF VARIATION CV, a statistic method, can reflect the variation of the observed values. The variation is becoming weaker with the decrease in CV. Also, CV reflects the stability of the sensors [35]. Namely, the larger CV, the worse stability [36]. Table 3 describes the CV of sixteen sensors. It can be VOLUME 8, 2020 FIGURE 2. Photograph of the practical ABS odor evaluation system and the interior detection device: (a) Photograph of the practical ABS odor evaluation system. Part A is an air generator with a filter, part B is a thermostatic heater with a sampling bottle, and part C is a detection device which involves a sensor array, a data acquisition card, two pumps, and three valves. (b) Photograph of the interior detection device. Part D is the data acquisition card and part E is the sensor array.
acknowledged that the variable coefficients of MQ-138, TGS-822, and TGS-861 were more than that of others. Thus, these three sensors were excluded due to the worse stability.

b: ANALYSIS OF VARIANCE
As a means of hypothesis test, ANOVA mainly makes statistical analysis for many relevant factors, and compares the random errors and the systematic errors. In this paper, each sensor was served as a factor, and the changes of voltage  values were checked through ANOVA. In general, the repeatability of the sensors is becoming better with the decrease in intra-class quadratic mean. The larger F value or the less p value means the preferable discrimination of the sensors. Table 4 describes the ANOVA of sixteen sensors.  Table 4, the mean square values of MQ-135 and MQ-138 were more than that of others, meaning that these two sensors with the poorer repeatability. Moreover, the p value of TGS-832 was much higher than that of others. With the above analysis, MQ-135, MQ-138, and TGS-832 were deleted from the sensor array.

c: PRINCIPAL COMPONENT ANALYSIS
As a multivariate statistical means, PCA is able to transform the initial data into linearly independent characteristic components, and retain most of the information of the original data. When the cumulative contribution rates in PCA are more than 80%, the coefficient of PC2 can indicate the selectivity of the sensors. Namely, the greater coefficient of PC2, the preferable selectivity. In this study, the cumulative contribution rates of the three principal components were 87.12%. As shown in Table 5, the coefficients of PC2 of MQ-135, MQ-138, TGS-831, TGS-2611, and 2m009 were less than that of others, which proved that these five sensors were provided with the worse selectivity.

C. EXPERIMENT APPROACHES
To begin with, the ambient temperature and humidity were at 20 ± 2 • C and 65 ± 5 % RH. Before measuring, open the equipment and the software interface to preheat the gas sensors for 1 h until the sensors reach steady voltage values. The formal operations are as follows: (1) Define the file names in the light of the experimental requirements (manufacturers xx, model number xx) and save the files in the specified path. Steps (1) -(4) were repeated to start next test.
It is well known that the response values of the sensors are sensitive to the environment. Consequently, the response values could be acquired by eliminating the baseline values. The response values were calculated by the following formula: where, X s is the measured value, and X b is the baseline value. To reduce the differences between the data and improve the convergence rate, the response values were supposed to be normalized. The min-max normalization formula is as follows: where, X max and X min are the maximum and minimum values of the response values, respectively. And the final normalization between [0, 1] was gained. As shown in Fig. 3, the sensors kept the initial values before contacting with the gases. After 5 s, the voltage response of each sensor raised fast and held a stable value at about 120 s. Therefore, the average value of each voltage response curve from 120 s to 139 s was obtained as the feature. There were four sorts of ABS, and four groups of 80 × 8 feature matrices could be gained.

A. GREY RELATION ANALYSIS
GRA is a method to analyze the correlation of various factors in the system, developed by Deng Julong [37]. The contribution of each factor to the system can be concluded based on the grey relational degree. The geometric similarity of the data sequence of each factor is studied to determine the correlation between the factors [38]. The closer the geometry is, the greater the correlation is. Because of good universality and reliability, GRA makes progresses in mathematics and economics [39]. The particular approaches are as follows: (1) Determine the referred sequence and the compared sequence. Suppose there are n evaluation indexes and m evaluation objects. Hence, the referred sequence could be described as: x 0 = {x 0 (k)|k = 1, 2, . . . , n}, and the compared sequence could be described as: (3) The correlation coefficient is obtained by solving the following problem: where ρ is the discrimination coefficient (0< ρ <1), usually recommended to be 0.5. (4) Finally, the correlation r i can be obtained from Equation (4):

B. EXTREME LEARNING MACHINE
ELM is a machine learning algorithm based on feedforward neural network, developed by Huang Guangbin [40]. Compared to the other networks, ELM sets the random weights (ω) and thresholds (b) of the hidden layer without adjustment, which reduces computation [41]. The output weights (β) are determined by the activation function and the amount of hidden layer neurons. Owe to the higher efficiency and the stronger generalization, ELM has an extensive application in classification and regression [42]. The detailed steps are as follows: The infinitely differentiable activation function of the hidden layer neuron is g(x), which can be described as: where β in is the connection weight of the i th hidden layer neuron and n th output layer neuron, β i = [β i1 , β i2 , . . . , β in ] T . w in is the connection weight of the i th hidden layer neuron and the n th input layer neuron, . b j is the threshold of the j th neuron on the hidden layer. (2) The purpose of SLFN is to minimize the output error.
Thus, W i and b i enable the result to approach to the t j : where H is the output matrix of the neuron on the hidden layer. (3) The output weights are able to be acquired by settling the least square solution of next formula: The solution is:β where H + is the Moore-Penrose generalized inverse of output matrix.

C. RANDOM FOREST
RF refers to an ensemble learning method that employs random growing trees to classify and predict [43], also termed as random decision trees. The combination of Bagging and random subspace addresses the weaknesses of single tree, which is suitable for the classification of high dimensional data [44]. When RF is applied to classification, every decision tree is performed to categorize the samples. The final category is determined by the category winning the most votes. As for regression, the final result is obtained by the mean of the predicted values of the decision trees [45]. Because of speediness and efficiency, this algorithm has progressed in data mining. The detailed descriptions are as follows: (1) Bootstrap is adopted to produce T training sets: S 1 , S 2 , . . . , S T , at random.

D. SUPPORT VECTOR MACHINE
As a model of machine learning, SVM makes striking progresses in pattern classification and regression analysis since proposed by Vapnik in 1964 [46]. The main theory is that a hyperplane is instituted to separate two types of data for the maximum interval. As for the nonlinear data, the nonlinear transformation is utilized to convert the input space into a high dimensional space, and the linear classification and optimal hyperplane could be completed [47]. The nonlinear transformation is accomplished by kernel function, such as linear function and polynomial function [48]. The following steps can express the concrete processing: (1) Set the training set: , −1}(i = 1, 2, . . . , l). The hyperplane is expressed as: Where ω is the weight vector, and b is the threshold. (2) In order to maximize the margin of each category, the optimization of hyperplane is solved under the following condition: (3) By introducing Lagrange multipliers a i and a * i , the solution to the optimal classification plane is converted into the corresponding dual problem: a i y i = 0 (9b) VOLUME 8, 2020 where c is the penalty parameter to measure the error of the model. The optimal solution is: (4) In this paper, the radial basis function is selected as the kernel function: where g is the kernel function parameter, concerned with the data distribution of feature space. The decision function can be described as:

A. QUANTIFICATION OF ODOR INTENSITY
According to GRA, the referred sequence is often the optimal sequence [49]. Compared with other sensors, MQ-2 and MQ-3 were equipped with striking response values. However, the response curves of MQ-2 were steadier than that of MQ-3 in the stable state. Therefore, the feature of MQ-2 was selected as the referred sequence. And the features of the other sensors were served as the compared sequences. As shown in Fig. 4, the correlation degree of each sensor was distinct from one another, which revealed the better contribution of the sensors in the array. The final quantified results are shown in Fig. 5. The quantified values of odor intensity were becoming greater with the increase in the weight of ABS, which conformed with the fact that the odor is becoming more intense with the increased concentration. Besides, the odor intensity of different ABS brands was evidently differentiated by the quantified numbers under the same weight. From the above, the odor intensity of ABS could be quantified by  the odor evaluation system. The odor of high heat resistant ABS was more intense than that of others, and the odor of flame-retardant ABS was the weakest.  To verify the above results, polypropylene and glass ABS, an extra brand of ABS, was measured by the odor evaluation system. As shown in Fig. 6, the quantified values of polypropylene and glass ABS were also getting greater with the increase in the weight of ABS. Besides, the odor intensity of polypropylene and glass ABS was distinguished from that of other brands when the samples were equipped with the same weight.

B. QUANTITATIVE REGRESSION OF ODOR INTENSITY
The regression models were constructed by ELM, RF, and SVM. For each brand, 60 groups of data were selected as the training set at random, and the remaining 20 groups of data were used as the test set. The input of the model was composed of the features of the odor and the output was made up of the quantified values of odor intensity. In order to describe the predicted performance of the models, the R 2 and the RMSE were introduced. The R 2 is termed as coefficient of determination, used to measure the fitting effect of the regression model. In general, the larger R 2 between [0, 1] means the better fitting effect. The RMSE, namely root mean square error, is served as an indicator to evaluate the deviation between the predicted values and observed values.

1) REGRESSION ANALYSIS BY ELM
The regression model was constructed by ELM, where the number of neurons was 20. As shown in Table 6, the R 2 > 0.92 in the training sets and the test sets, and the difference of R 2 between the training sets and the corresponding test sets was less than 0.02. Therefore, ELM possessed good universality and generalization for four kinds of ABS. Besides, the quantified values of heat-resistant ABS were predicted well in the training sets (R 2 > 0.98) and the test sets (R 2 > 0.97). The fitting effect of the predicted values and the quantified values in the test sets is shown in Fig. 7. VOLUME 8, 2020

2) REGRESSION ANALYSIS BY RF
In RF, the number of the decision trees was 500, and the number of attributes in the split attribute set was 5. Table 7 shows the regression results of four kinds of ABS. It can be seen that the training sets with R 2 > 0.91 and the test sets with R 2 > 0.89. And the difference of R 2 between the training sets and the corresponding test sets in ELM was less than that in RF, which proved that ELM was better at applicability and generalization. Among four kinds of ABS, the quantified values of polypropylene and glass ABS were predicted better, with R 2 > 0.99 in the training sets and R 2 > 0.98 in the test sets. Fig. 8 shows the fitting effect of the predicted values and the quantified values in the test sets.

3) REGRESSION ANALYSIS BY SVM
The Cuckoo Search (CS) was employed to search the best parameters (c and g). And the mean square error was utilized as the fitness function. As shown in Table 8, SVM had a preferable presentation in the training sets (R 2 > 0.98) but not the test sets (0.88 < R 2 < 0.97). The difference of R 2 between the training sets and the corresponding test sets in SVM was greater than that in the other models, which certified that SVM was provided with inferior fitting. The quantified values of polypropylene and glass ABS were also predicted well by SVM, with R 2 > 0.99 in the training sets and R 2 > 0.96 in the test sets. The fitting effect of the predicted values and the quantified values in the test sets is shown in Fig. 9.

V. CONCLUSION
In this paper, an odor evaluation system was developed to detect the odor intensity of four kinds of ABS. Eight sensors were selected to compose an array. Then, the data preprocessing and the feature extraction were applied to deal with the data of smells. Finally, the correlations and the weights of different sensors were obtained to quantify the odor intensity. According to the quantified consequences, ELM, RF, and SVM were introduced to construct the regression models. The particular conclusions are as follows: (1) Design of the ABS odor evaluation system: the system included a sampling device, a sensor array, and an analysis software. Eight sensors were screened by CV, ANOVA, and PCA to construct an array with stability, repeatability, and selectivity. This study indicated that the odor evaluation system could perceive the smells of four kinds of ABS and quantify their severity, offering a novel technique for the evaluation of the odor intensity. Thus, this system was likely to substitute human evaluation in the field of automobiles. JINGJING LIU was born in Jilin, China, in 1986. She received the Ph.D. degree in agricultural electrification and automation from Jilin University. Since 2013, she has been an Associate Professor and a master's Tutor with the School of Automation Engineering, Northeast Electric Power University. She is the author of over 30 articles and holds four patents. Her main research interests include sensor array construction, multisensor data fusion, and evaluation of sensory information. VOLUME 8, 2020