Information: This paper is based on the publication “Using artificial intelligence (AI) in electronic nose system (E-Nose) development for air quality detection in shopping centres” presented at the 18th Healthy Buildings Europe Conference 2023, Aachen, Germany.


Mahmoud ElMokadem
Samy Louca
Ali Abdelrahman
Kai Rewitz
Dirk Müller
Institute for Energy Efficient Buildings and Indoor Climate (EBC), E.ON Energy Research Center, RWTH Aachen University, Aachen, Germany.


Abstract. With the increasing awareness concerning indoor air quality (IAQ) and energy efficiency, the development of IAQ sensor systems is becoming more relevant, for example to enable demand-controlled ventilation (DCV). Currently, DCV in most indoor environments is based on CO₂ concentration, which is a good indicator for human emissions. For shopping centres (SC), DCV rarely considers volatile organic compound (VOC) emissions from materials, which could significantly affect customer health and consumer experience. VOC multi-gas sensor systems (VOC MGSS)can help to detect and interpret these gases. Due to the high number of sensors and influencing factors such as temperature, relative humidity or type of odour, choosing the best model approach for evaluating IAQ is not trivial. Therefore, we compare linear regression with support vector regression, artificial neural networks and random forest models for the calibration of a VOC-MGSS to predict perceived intensity (PI) for six different SC products. Subjective test data from a trained subject panel and objective test data from the VOC-MGSS are used to train the different model types for the overall dataset and product-specific (PS) datasets.


Keywords. Indoor air quality, total volatile organic compounds, artificial intelligence, E-Nose, linear regression, support vector regression, artificial neural networks, random forest regression


1. Introduction

Several studies show the impact of indoor air quality (IAQ) among others on human health, occupant wellbeing, comfort and productivity and consequently the need for efficient ventilation strategies [1]. One way of achieving high IAQ with low energy requirements is to use demand-controlled ventilation (DCV). Currently, DCV is mostly based on CO₂ concentration, which is a good indicator of human emissions. However, in SC releasing potentially hazardous VOC from seemingly harmless sources is proven to cause high potential risk to human health, causing sick-building syndrome (SBS) or building-related illnesses [2]. Thus, identification of VOC-inducing products and correlation to subjective evaluation data are essential for a healthy and comfortable environment in SC.

Compared to expensive and complex to use gas chromatography-mass spectrometry, VOC-MGSS are cheap and convenient testing methods to detect VOCs in indoor environments [3, 4]. For such VOC-MGSS, different sensor types can be used. An often-used sensor type is the metal oxide semiconductor (MOx) sensor. Among other benefits being high sensitivity to low gas concentrations, long service life and a low price, MOx sensors are potentially suitable for use in DCV systems. Examples of VOC-MGSS with MOx sensors are available for research use, such as KAMINA from the Karlsruher Institute of Technology (KIT) [5] and for commercial use, such as i-Pen, PEN2, and PEN3 from Airsense Analytics [3] to assess air quality in indoor areas.

Since most of these devices have data from more than one sensor, implementing an artificial intelligence framework can be used to correlate the output of all sensors with target variables, for example to distinguish between different odours or to predict perceived intensities. Earlier research combines artificial intelligence methods and VOC-MGSS technology successfully. Soh et al applied artificial neural networks (ANNs) for herbs recognition [7]. The accuracy of the model reaches 90% in the best case. Wang et al. used a random forest (RF) model for the concentration prediction of binary mixed gases, with the model having an R² score of 0.98 [8].

For the assessment of SC air quality, we already developed an VOC-MGSS mainly based on MOx sensors [6]. Different SC product types were tested with the VOC-MGSS and a trained subject panel to evaluate the PI. This testing aimed to find a correlation between objective and subjective data. In this previous work, only one sensor within the VOC-MGSS system was investigated and linear regression analysis was used [6]. For improved analysis, our goal is to investigate different AI frameworks for this application and find the most sufficient model that will enable the VOC-MGSS system to apply all its sensors when calculating the PI.

2. Methods

2.1 Objective and Subjective tests

The preceding study was conducted in the Air Quality Laboratory of the EBC. The SC product groups were divided into six categories: books, shoes, clothing, coffee and two perfume variants (100% and 40% diluted). These were filled into so-called emission chambers, shown in Figure 1. The chambers are connected to an HVAC system to ensure constant temperature and humidity. Inside each chamber beside the product, the VOC-MGSS was positioned to detect the VOC emissions from the products. A more detailed description can be found in [6].

Figure 1. Emission chambers in the Air Quality Laboratory at EBC.

The essential component within the developed VOC-MGSS system is a MOx sensor array of the MQ sensor series, which can detect a range of VOCs. The MQ sensors were manually pre-calibrated by adjustment of the potentiometer and post-calibrated to temperature and humidity with respect to the log curves available in the datasheets. All applied sensors in this system are outlined in Table 1. The VOC-MGSS prototype, including optimized housing, can be seen in Figure 2.

Figure 2. VOC-MGSS prototype of EBC, RWTH Aachen University.


Table 1. Sensors and detectable gases within VOC-MGSS system.


Detectable gases



Hydrogen, methane, propane,
i-butane, LPG , alcohol, smoke



Alcohol, ethanol, smoke



Carbon monoxide



Carbon monoxide, flammable gasses



Ammonia, nitrogen oxide, alcohol, benzene, smoke, carbon dioxide



Benzene, toluene, alcohol, formaldehyde gas, hydrogen



Industrial sensor for TVOCs



For subjective evaluation of the emission intensity from the SC products, a trained panel consisting of 17 participants aided with an acetone test stand was used according to DIN ISO 16000-28. The conducted tests ran with a comparative acetone scale ranging from 0 to 28 PI.

The tests were carried out in respect with variations in temperature (20, 23, 27 °C), relative humidity (30, 50 %) and the air change rate (1, 1.7, 2.05, 3.15, 4.1, 5.1 h−1). In total, 36 evaluations per product and test subject were recorded and stored in a dataset.

2.2 AI Model development procedure

To find a suitable model to predict the perceived intensity depending on the sensor data, different AI models were created and assessed on the MATLAB Regression Learner App. This GUI allows for training, validation, and tuning of models. In our context, sensor signals in Table 1 are categorized as input data, also called predictors, while the corresponding perceived intensity is the known response. External factors are excluded from predictor selection as an ideal sensor calibration is assumed.

Despite all model types operating differently, the developing approach for each model type is similar, which can be mainly divided into a training and testing process. The model developing procedure for the AI models (AI MDP) is shown in Figure 3.


Figure 3. AI model development procedure (AI MDP).

2.2.1 Dataset pre-processing

After a random permutation of the dataset it is separated into training and testing datasets by using 80/20 split [12]. The following step is setting specific model parameters and training of the model with the training dataset. Cross-validation values are calculated to determine the training success[13]. Finally, the trained models are tested to evaluate their performance. To investigate if a product specific (PS) model development leads to improvements in the predictions of the different model types, AI MDP is applied for overall data and each product separately. This leads to six additional product specific models for each model type.

2.2.2 AI model setup

AI models significantly differ in behaviour, algorithmic concept and general bias/variance relevance for different model types in traditional AI problem-solving, machine learning, and deep learning applications [14]. For this, four model types were tested to evaluate their suitability, see Figure 3.

Linear regression finds the best parameter fitting for a regression function with the determination of RMSE. Despite its simplicity, it is subject to underfitting and is affected by outliers [15].




Equation (1) shows the applied linear regression function. In this study, y is the predicted perceived intensity, βn are the fitted coefficients and xnare the predictors.

Support vector machine (SVM) regression involves the addition of support vectors and kernel functions to estimate a target function within a set constraint (ε) and eliminate outliers (ϑ) from its analysis. The kernel functions can include higher polynomials in finding an optimal hyperplane. In our case, linear and quadratic models are considered. However, the disadvantage is in optimizing the constraint (ε), which can cause a high error tolerance [16]. Figure 4 visualizes the regression process of a support vector machine with MQ-2 sensor data from this study.

Figure 4. Sample prediction of perceived intensity using support vector regression with MQ-2 sensor output.

As our system includes multiple MQ sensors more complex models in ANN can run those through numbered sets of layers and nodes[17]. In this study, two network configurations of one layer and 25 nodes and two layers with ten nodes each are used. The main advantage of ANNs is in detecting all possible interaction between predictors. Despite this, ANNs can easily over-fit, and require great computational resources [18, 19]. Figure 5 visualizes an example of an ANN with sensor signals as inputs, the PI prediction as its output.

Encompassing the same idea with a different execution is random forest (RF) algorithms. Here, the general aim is within constructing a number of decision trees for training [20]. These trees operate with leaves representing data values and applying binary thinking. The leaf size (LS) is related to the number of observations considered in each data split. Thus, RF models with small (4) and large (36) LS were applied in our case. The benefits of RF models are their robustness against outliers. In contrast to this, they are affected by small changes in data [15]. An example of a random forest model is seen in Figure 6.

Figure 5. Simplified artificial neural network (ANN) structure in application with VOC-MGSS input & output data.

Figure 6. Simplified visualization of random forest algorithms.

2.2.3 Model testing

To build models that accurately predict the PI from objective data, evaluation of the suitability of each model to the dataset is required. RMSE shows how much the model predictions deviate from the subjective test data. [21]. Equation (2) shows the definition for the RMSE. In this study, yi is PI of the subject test, m(AI) is the prediction of each model andn represents the number of data values.



3. Results

3.1 Training and testing of the overall model

Table 2 presents the RMSE for training and testing for the whole dataset. Figure 7 shows no major differences in the distribution of the data between the subjective test (red boxplot) and LR, ANN, RF and Quadratic SVM models (grey boxplot), only a small difference is observed for Linear SVM. Table 2 shows the best performance for the RF model with a leaf size of 4 and 36, achieving testing RMSE of 0.07 and 0.75 PI respectively. The ANN models (1x25 and 2x10) also show a good performance, achieving testing RMSE of 0.85 and 0.86 PI.

Table 2. Resulting training and testing RMSE for the overall models.

AI model

RMSE in PI Training

RMSE in PI Testing

Linear Regression



Linear SVM regression



Quadratic SVM regression



ANN (2x10)



ANN (1x25)



RF (leaf size: 4)



RF (leaf size: 36)




Figure 7. Distribution of subjective test data and prediction of the overall models.

3.2 Testing of the product specific models

Figure 8 shows the RMSE for the overall models (red bars) and the RMSE for each PS model (grey bar). The best performance of the PS models has RF-LS 4 scored a maximum RMSE of 0.09 PI, which is close to the RMSE of 0.07 PI for the overall model. Increasing the LS to 36 results in an increased maximum RMSE of 0.8 PI for the PS models. The second ranking is for the ANN models (2x10 and 1x25) achieving a maximum RMSE of 0.33 PI and 0.43 PI. These are lower than the RMSE of 0.86 PI and 0.85 PI for the overall models.

Figure 8. Comparison of the RMSE in PI for the overall and product specific models for testing.


Figure 8 shows that for LR, SVM, and ANN the PS RMSE values are smaller than overall RMSE values. For RF models the RMSE of overall model and PS models are more similar. The results indicate that the PS models should be used for their respective stores in the SC to predict PI rather than the overall model.

4. Conclusion

The results show the potential of VOC sensors for IAQ assessment in SC. Subjective and objective tests were conducted to validate the application of this technology with SC products. The use of AI models enables a correlation between the VOC-MGSS sensor signals and subjective values. To achieve this, an analysis of five AI models is conducted. The models were trained and tested with both overall and PS datasets. For model evaluation, RMSE metrics were applied. The best performance shows RF-approach-LS 4, achieving for the overall and SP dataset 0.07 PI and max. 0.09 PI. However, there may be an overfitting for this model [20]. For all model types except RF models, the results show a decreased RSME for product specific models compared to the overall models. The ANN models (2x10) and (1x25) achieved RMSE of 0.86 and 0.85 PI. This makes them also applicable for this dataset. Overall, further analyses with more subjective and objective data, especially validation data from field tests, are needed to bring such VOC-MGSS system into practice use. Moreover, a correlation of PI and acceptance, is needed for implementing these models in DCV concepts.

5. Acknowledgements

Grateful acknowledgment is made for financial support by the Federal Ministry for Economic Affairs and Climate Action (BMWK) regarding the project “Optimization of modularly planned shopping centres supported by energy data”, promotional reference 03ET1502.

6. References

[1]     Mohamed A-MO, Paleologos EK, Howari FM, editors. Pollution assessment for sustainable practices in applied sciences and engineering. Kidlington: Butterworth-Heinemann; 2021.

[2]     Shang Y, Li B, Baldwin AN, Ding Y, Yu W, Cheng L. Investigation of indoor air quality in shopping malls during summer in Western China using subjective survey and field measurement. Building and Environment. 2016;108:1–11. doi:10.1016/j.buildenv.2016.08.012.

[3]     Wilson AD, Baietto M. Applications and advances in electronic-nose technologies. Sensors (Basel). 2009;9:5099–148.

[4]     Wilson AD. Review of Electronic-nose Technologies and Algorithms to Detect Hazardous Chemicals in the Environment. Procedia Technology. 2012;1:453–63. doi:10.1016/j.protcy.2012.02.101.

[5]     Goschnick J, Harms M. Anwendungsstudie des Karlsruher Gassensor-Mikroarrays für die schnelle Brandgasanalytik; 2000.

[6]     El-Mokadem M, Rewitz K, Müller D. Influence of temperature and relative humidity on subjective and objective air quality data in shopping centers 2022. doi:10.34641/CLIMA.2022.418.

[7]     Soh AC, Chow KK, Yusuf UKM, Ishak AJ, Hassan MK, Khamis S. Development Of Neural Network-Based Electronic Nose For Herbs Recognition. International Journal on Smart Sensing and Intelligent Systems. 2022;7:584–609. doi:10.21307/ijssis-2017-671.

[8]     Wang T, Wu Y, Zhang Y, Lv W, Chen X, Yang Z. Concentration prediction of binary mixed gases based on random forest algorithm in the electronic nose system. In: Li Q, editor. Thirteenth International Conference on Signal Processing Systems (ICSPS 2021); 11/12/2021 - 11/15/2021; Shanghai, China: SPIE / International Society for Optical Engineering; 2022. p. 42. doi:10.1117/12.2631552.

[9]     Liao Y-H, Shih C-H, Abbod MF, Shieh J-S, Hsiao Y-J. Development of an E-nose system using machine learning methods to predict ventilator-associated pneumonia. Microsyst Technol. 2022;28:341–51. doi:10.1007/s00542-020-04782-0.

[10]   Patel K, Raina S, Gupta S. Artificial Intelligence and its Models. 2020.

[11]   Mathworks. Machine learning technique for building predictive models from known input and response data. 2021.

[12]   Gholamy A, Kreinovich V, Kosheleva O. Why 70/30 or 80/20 Relation Between Training and Testing Sets: A Pedagogical Explanation; 2018.

[13]   Mathworks. Cross-Validation: Assess and improve predictive performance of models. 2021.

[14]   Jordan MI, Mitchell TM. Machine learning: Trends, perspectives, and prospects. Science. 2015;349:255–60. doi:10.1126/science.aaa8415.

[15]   Shanmugasundar G, Vanitha M, Čep R, Kumar V, Kalita K, Ramachandran M. A Comparative Study of Linear, Random Forest and AdaBoost Regressions for Modeling Non-Traditional Machining. Processes. 2021;9:2015. doi:10.3390/pr9112015.

[16]   Zhang F, O’Donnell LJ, editors. Machine Learning: Methods and applications to brain disorders: Support vector regression: Elsevier; 2020.

[17]   Choi RY, Coyner AS, Kalpathy-Cramer J, Chiang MF, Campbell JP. Introduction to Machine Learning, Neural Networks, and Deep Learning. Transl Vis Sci Technol. 2020;9:14.

[18]   Tu JV. Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol. 1996;49:1225–31. doi:10.1016/s0895-4356(96)00002-9.

[19]   Di Franco G, Santurro M. Machine learning, artificial neural networks and social research. Qual Quant. 2021;55:1007–25. doi:10.1007/s11135-020-01037-y.

[20]   Breiman L. Random Forests. Machine Learning. 2001;45:5–32.

[21]   DataTechNotes. Regression Model Accuracy. 2019.

[22]   Berglund LG. Comfort and Humdity. ASHRAE Journal. 1998:35–41.

Mahmoud ElMokadem, Samy Louca, Ali Abdelrahman, Kai Rewitz, Dirk MüllerPages 28 - 32

Stay Informed

Follow us on social media accounts to stay up to date with REHVA actualities


0 product in cart.products in cart.