CET-vol95


CHEMICAL ENGINEERING TRANSACTIONS 

VOL. 95, 2022 

A publication of 

The Italian Association 
of Chemical Engineering 
Online at www.cetjournal.it 

Guest Editors: Selena Sironi, Laura Capelli 
Copyright © 2022, AIDIC Servizi S.r.l. 
ISBN 978-88-95608-94-5; ISSN 2283-9216 

Electronic Nose for Real-time Monitoring of Odour Emissions 
at a Wastewater Treatment Plant 

Stefano Prudenza, Alessandra Panzitta, Carmen Bax*, Laura Capelli 
Politecnico di Milano, Departement of Chemistry, Materials and Chemical Engineering, Piazza Leonardo da Vinci, 32, 20133 
Milano, Italy. 
carmen.bax@polimi.it 

This paper describes the procedure adopted for implementing a monitoring system based on an electronic nose 
(EN) to continuously monitor odour emissions from a wastewater treatment plant (WWTP), with the purpose of 
identifying odour peaks related to the incoming wastewaters. The paper focuses on the methodology related to 
the instrument training, the implementation of a suitable model to classify the odour peaks, and then the 
validation procedure. In the specific case, the EN was installed at the arrival tank of the plant, which is 
characterized by anomalously intense odours and high variability of the incoming wastewater, with the purpose 
of identifying the origin of the odour peaks. To do this, the EN was equipped with automatic sampling systems 
to collect both the liquid effluent and gaseous samples at the arrival tank for further olfactometric and chemical 
characterizations. The paper limits its focus on the illustration of the EN training to identify the anomalous odour 
peaks related to unpredictable changes in the incoming wastewater, and the validation of the implemented 
model based on principal component analysis and support vector machine. Results achieved show that the EN 
can be effectively used for process control: the alarm set on EN signals proved effective in detecting alterations 
of the incoming effluent potentially responsible for odour events in the surroundings of the plant, thereby allowing 
plant managers prompt intervention to limit odour impacts.  

1. Introduction
Electronic Nose (EN) can serve as useful tool for monitoring plant emissions at their fenceline or receptors in 
the surroundings, since they are capable to provide a real-time qualitative and quantitative characterization of 
odours in ambient air. An emerging application of this technology is emission monitoring aimed at process 
control (Collins and Moy 1995, Martı́, Busto et al. 2005, Majchrzak, Wojnowski et al. 2018) . Indeed, the 
continuous characterization of odorous emission can be useful to verify that the plant is working under regular 
operating conditions, and for the real-time identification of malfunctioning and/or anomalies that could cause 
odours nuisance nearby the plant. This paper proposes the use of an EN for the continuous monitoring of 
odorous emissions from a WasteWater Treatment Plant (WWTP). The WWTP under investigation receives both 
civil and industrial wastewaters from over 40 industries of different types (e.g., chemical, tanneries, paper mill, 
printing, food, car washes etc.). Civil wastewaters account for about 60-70% of the total incoming flowrate (i.e., 
ca. 22’000 m3/day), while industrial wastewaters account for the remaining 30-40%. Due to the frequent reports 
of odour nuisance from the citizens living in the proximity of the plant, in agreement with the regional guidelines 
on odour emission management, several olfactometric campaigns were carried out at the WWTP emission 
sources with the purpose of determining odour concentrations and odour emission rates. This preliminary work 
allowed identifying the sources with the highest odour emissions and evaluating the odour impact associated to 
the plant in the surrounding territories. As expected, based on experimental evidences collected so far, the 
arrival tank turned out to be the most problematic source of the plant. One peculiarity of this plant is the 
considerable variability of the odour concentration values measured at the arrival tank, which ranges from 2’000 
to 120’000 ouE/m3. This variability is not related to the meteorological conditions, but is more likely linked to the 
arrival of a particularly odorous and discontinuous discharge originating those malodorous events. However, 
the origin of this particularly odorous discharge has not yet been identified. Since these particular odorous 

 DOI: 10.3303/CET2295029 

Paper Received: 12 March 2022; Revised: 26 May 2022; Accepted: 27 June 2022 
Please cite this article as: Prudenza S., Panzitta A., Bax C., Capelli L., 2022, Electronic Nose for Real-time Monitoring of Odour Emissions at a 
Wastewater Treatment Plant, Chemical Engineering Transactions, 95, 169-174  DOI:10.3303/CET2295029 

169



conditions occur in an unpredictable manner, it was decided to implement a monitoring system capable to 
provide a real-time measurement of the odours at the arrival tank and to detect anomalous conditions.  
Once the occurrence of an odour peak is detected, the system shall automatically activate a sampler to withdraw 
both air and liquid samples to be analysed by dynamic olfactometry and GC-MS for the identification of the 
chemical substances related to the odorous wastewater discharge causing the odour peaks. The chemical 
speciation of the substances contained in the liquid and in its headspace, compared with the respective olfactory 
threshold values, is used to identify the chemicals most responsible for the odour emissions. This paper focuses 
on the experimental procedure involved for the training of an EN for the implementation of such monitoring 
system and the evaluation of its performance. More in detail, the EN, equipped with an automatic gas sampling 
system, has been installed at the arrival tank of the WWTP. The training lasted about 1 month and involved the 
analysis of samples collected at the arrival tank under different meteorological and operating conditions of the 
plant. EN signals relevant to the training period were combined with the odour concentration values measured 
by dynamic olfactometry and with reports of presence/absence of odour by the citizens living nearby, to 
implement classification models aimed at identifying a Normal Operating Region (NOR) (Leung and Romagnoli 
2002), representative of moderate presence of odour on the arrival tank not causing nuisance in surrounding 
territories. This paper also reports the results of the model validation. In this phase, gas samples have been 
randomly collected at the plant in order to verify the peformance of the model developed.  

2. Material and methods 
2.1 Electronic nose and sampling system 

The EN used in this project is the WT1 outdoor monitoring system commercialized by Rubix. This instrument is 
characterized by a fast response time, returning the sensors’ signal every 10 seconds, and is therefore suitable 
for real-time monitoring purposes. It is equipped with 4 metal oxide sensors (MOS), characterized by a high 
sensitivity to Volatile Organic Compounds (VOC), 3 electrochemical sensors of for the detection of hydrogen 
sulphide (H2S), formaldehyde (CH2O) and ammonia (NH3), and a photoionization detector (PID) calibrated in 
isobutylene for VOC. The MOS sensors return as response the value of the sensor resistance in Ohms, while 
the electrochemical sensors and the PID provide the concentration of the compounds in ppm. The instrument 
comprises also sensors for measuring the temperature and relative humidity of the external environment.  
Together with the EN, also a gas automatic sampler has been installed at the arrival tank of the WWTP. It 
consists of an airtight suitcase of 25 liters produced by Scentroid (VC20) equipped with a membrane vacuum 
pump and Nalophan bags for sample collection. The sampler can be activated manually or automatically by the 
EN when the set alarm threshold is exceeded.    

2.2 Training of the Electronic Nose 

In this study, the training phase lasted for about 30 days during which the EN was installed inside the shed of 
the arrival tank to acquire data representative of the odour conditions at the plant inlet. Due to the extreme 
environmental conditions inside the arrival tank shed (high relative humidity and high levels of hydrogen 
sulphide), in order to avoid excessive sensors’ deterioration, a dilution system was applied at the inlet of the EN. 
The system consists of an Y joint attached to the inlet of the nose, to which 2 Teflon tubes were connected, one 
taking the air from the inside of the shed and one from the ambient outside. Exploiting the sucking capability of 
the WT1 and by tailoring the Teflon tube length a 1:1 dilution of the flow coming from the arrival tank with ambient 
air has been reached.  The training phase consisted in the implementation of a model able to correlate the EN 
sensors’ responses with the level of the odour concentration on the arrival tank. The training was aimed at the 
definition of a “reference” condition on the arrival tank of the WWTP representative of moderate odour 
concentration, unlikely to cause nuisance to the citizens living nearby the plant. For this purpose, the training 
phase involved the collection of gas samples inside the arrival tank shed and their characterization by dynamic 
olfactometry. In particular, 18 samples were collected on 10 different days under different meteorological and 
operating conditions of the plant (e.g., different incoming wastewater flowrate, active or inactive recirculation of 
sludges) in order to include in the training dataset the variability of the source. Once the data were collected, a 
specific procedure, illustrated in Figure 1, has been developed for their elaboration The sensors response is 
initially averaged on 5 minutes, to reduce the dataset dimensionality, and ease data elaboration and 
interpretation. It has been verified that this operation does not cause loss of useful information for the 
differentiation of the odour level operated by subsequent processing. Then, a principal component analysis 
(PCA) (Abdi and Williams 2010) was applied to further reduce the data dimensions, allowing for a better 
visualization and investigation of their structure. By looking at the PCA scores distribution and combining the 
information about the odour concentration detected on the arrival tank and citizen reports, a preliminary 
“reference” region on the score plot has been drawn.  

170



Finally, in order to define a more rigorous NOR representative of moderate odour conditions at the inlet of the 
plant, the PCA scores were used as inputs of Support Vector Machine (SVM) algorithm (Sun 2014), in order to 
obtain an optimal decision boundary according to data distribution and the information relevant to the odour 
concentration. 
 

 
Figure 1. Block diagram of the data elaboration for the development and validation of the model. 

2.3 Validation 

Once the alarm model was developed and implemented on the EN, the validation and monitoring phase started. 
The data registered by the EN in this phase follows the same data processing pathway implemented for the 
training. The sensors response is averaged over 5 minutes and then the data are projected onto the PCA model 
developed on training data. Then, the PCA score obtained on monitoring data were used as inputs for the SVM 
algorithm to determine if they fall inside or outside the defined NOR. In case of projection outside the NOR 
boundaries, the EN reports an alarm and activates the automatic samplers for collecting gaseous and liquid 
samples to be further analysed by dynamic olfactometry, with the purpose of evaluating if their odour 
concentration was consistent with the alarm exceedance. In order to validate the model, validation samples 
were collected both in case of alarm threshold exceedance and under normal conditions. A total of 10 validation 
samples was considered. 

3. Results 
3.1 Training 

3.1.1 Preliminary considerations 
The dataset obtained at the end of the training phase consists of 246570 observations from which 6 features 
each have been extracted (i.e., the signals of 2 MOS, 3 electrochemical sensors and 1 PID sensor). Here only 
2 out of the 4 MOS sensors have been considered (i.e. one more sensitive to amine compounds and one more 
sensitive to aldehydic compounds), since the other were more subjected to the variation of relative humidity and 
temperature registered on the arrival tanks and did not provide useful information. These data have been 
processed by PCA, and the results are reported in Figure 2. For this case study, 2 principal components, 
describing 72.7% of the dataset variance, have been considered. More in detail, Figure 2 A reports the PCA 
score plot, where the projection of the data into the new reference system is reported. The plot also reports the 
odour concentration in ouE/m3 of the samples taken from the arrival tank assessed by dynamic olfactometry. 
Based on visual investigation of the PCA score plot, the PC1 direction expresses a clear correlation with the 
odour concentration of samples: decreasing PC1 values correspond to increasing odour concentration.  
Figure 2 B reports the PCA loading plot, which provides information about the correlation among variables and 
their contribution to the dispersion of points in the PCA score plot. From its visual investigation it can be deduced 
that the left region of the plot accounts for odorous conditions at the WWTP. In particular, the high values of 
electrochemical sensors, corresponding to an increase of the detected concentration of H2S, NH3, CH2O and 

171



VOC, contribute to the positioning of observations in the left portion of the PCA score. Conversely, the MOS 
sensors’ loadings point in the opposite direction. Therefore, in the left portion of the plot, low resistance values 
are expected, in line with the MOS n-type behaviour when exposed to gas (i.e., their electrical resistance 
decreases).  

 
Moreover, the loading plot also enables determining correlations between the sensors’ signals. In this case, the 
MOS sensors are mainly correlated with the H2S, CH2O and PID sensors since the angle between them is 
narrower than 90°, the one indicative of no correlation. Conversely, there is a negligible correlation with the NH3 
sensor, because their angle is close to 90°. This is probably due to the lower levels of ammonia registered on 
the arrival tank, varying among 0.3 and 1.5 ppm. Combining the points distribution on the score plot and the 
information coming from the loading plot, and looking at the scores distribution on the PC2, three main regions 
of odours can be identified: one in the top-left part, one in the middle-left and one in the bottom-left part of the 
score plot. These 3 regions are relevant to different odour conditions at the arrival tank, associated to high levels 
of VOC, H2S/CH2O  and NH3 respectively. It is clear that this doesn’t mean that the presence of odours is only 
attributable to these compounds, but this information gives an estimation of the class of molecules most 
responsible for the specific odour peak.  
 
3.1.2 Development of the model 
Based on the scores distribution, the information obtained from the loading plot, the results of the olfactometric 
analysis and the citizen reports, a preliminary NOR, representative of moderate odour concentrations on the 
arrival tank, has been defined on the score plot. The citizen reports have been collected through the app 
Odorcollect and have been used for comparing the odour concentration at the arrival tank of the plant provided 
by the IOMS and the possible nuisance perceived by the citizen living nearby. In this way it has been possible 
to identify a score plot “region” representative of such condition that causes nuisance in the citizens. For the 
definition of this preliminary NOR only the points that satisfied the following conditions have been selectd: PC1>0 
˄ ∀ PC2, -2.5<PC1<0 ˄  -3<PC2<2. Once selected, they have been used to train a one-classification SVM model 
(Schölkopf, Platt et al. 2001), in order to determine a more accurate confidence region though radial kernel. This 
algorithm defines in an unsupervised manner a hyperplane/function around the data used to train it, that enables 
their circumscription.  

Figure 2. Scores (A) and Loadings (B) plot of the PCA on the training data. The numbers displayed in the 
score plot indicates the odour concentration of the gas samples taken from the arrival tank. 

172



The algorithm will classify the new data in a binary way, as belonging to the circumscribed region defined by the 
hyperplane or not. It is important to underline how, for this one-classification SVM, the choice of correct reference 
data is a key step that can have a great impact on the subsequent results. In this case, the use of different 
indicators for the identification of the reference condition on the arrival tank, such as EN signals, dynamic 
olfactometry results and citizen reports, provide a stronger base for the NOR definition. Support vectors define 
the margins of the hyperplanes, that can vary in shape and dimensions, depending on the input parameters 
provided by the user on the base of considerations on the data distribution and requirements on the 
classification. In particular, two parameters are required: the “nu” parameter that controls the sensitivity of the 
support vectors and should be tuned to the approximate ratio of outliers in the data, and the “gamma” parameter 
that defines the shape of the radial kernel. For this case study, the “nu” and “gamma” parameters were set at 
0.0001 and 0.03, respectively. With the practical purpose of introducing also an intermediate alarm region, 
representative of anomalous conditions that cannot be assimilated to the NOR region, but not so troubling in 
terms of odour concentration, a second area has been identified. This region is defined by PCA scores satisfying 
the following requirements: PC1>0 ˄ ∀ PC2, -4<PC1<0 ˄ -4<PC2<3.  

 
Figure 3. PCA score plot illustrating NOR and intermediate alarm region obtained with one-class SVM. 
 
On this data the SVM has been implemented to define the boundaries of the intermediate region, using as input 
parameters nu=0.0001 and gamma=0.03. The confidence regions obtained are reported in Figure 3.  
From the results obtained it is possible to observe that the NOR defined includes samples with odour 
concentrations between 160 ouE/m3 (with confidence interval (C.I.) 70 - 290 ouE/m3) and 2900 ouE/m3 (C.I. 1300 
- 5300 ouE/m3). The intermediate region comprises samples with odour concentrations between 5800 ouE/m3 
(C.I. 2500 - 11000 ouE/m3) and 6900 ouE/m3 (C.I. 3000-13000 ouE/m3) approximately, while samples with higher 
concentrations fall into the alarm zone.  

3.1.3 Validation and monitoring 
During the monitoring phase, the data are continuously acquired by the EN and processed by the models 
developed on the training data. With the purpose of evaluating the capability of the EN to detect the presence 
of anomalous odour conditions, a validation was carried out. For such purpose, 10 independent samples, taken 
on 5 different days and representative of both reference and anomalous odour conditions at the arrival tank, 
were collected. They were analysed with dynamic olfactometry in order to determine their odour concentration 
and verify the correct prediction of the developed model compared to the EN predictions.  
Figure 4 illustrates the graphical results of this validation phase. As can be noted, all the 4 samples with odour 
concentration below 4’900 ouE/m3 fall correctly within the NOR region, leading to a 100% recall for this class. 
The 2 samples with a concentration of 9’200 ouE/m3 fall outside the NOR: one into the intermediate (i.e., orange 
region in Figure 4) and the other one into the alarm region (i.e., red region in Figure 4), leading to a recall of 
50%. On the contrary, among the 4 samples with odour concentration higher than 11’000 ouE/m3, 3 correctly 
fall into the alarm region, while 1 falls inside of the NOR, bringing the recall of such class equal to 75%. Globally, 
the implemented model achieves a balanced accuracy of 75% (CI95% 55 – 99 %). Thus, the model proved to 
be suitable for the discrimination between conditions representative of normal functioning of the plant 
characterized by odorous emissions not causing odour events in the surroundings and anomalous odour peaks. 
Moreover, the alarm defined turned out to be effective in signalling the exceedance of the critical odour 
concentration and activating automatic sampling systems with the purpose of identifying the causes of the odour 
event detected.  

173



Misclassifications recorded could be associated to the wide confidence interval of dynamic olfactometry, making 
that odour concentrations of 11’000 and 9’200 ouE/m3 are substantially very similar. It is also worthy to highlight 
that this paper presents a preliminary model developed not for accurately quantifying the odour concentration, 
which is the objective of future developments of the work currently ongoing, but to provide an estimate of the 
odour level at the arrival tank to be used for activating automatic sampling of gas and liquid for further analysis.  

 
Figure 4. The PCA score plot with the NOR and intermediate region obtained with the One class SVM on 
external validation data. 

4. Conclusions 
This paper describes the EN monitoring of odorous emissions from the arrival tank of a WWTP. The EN was 
trained to signal the detection of anomalous malodours by implementing specific multivariate model, based on 
PCA and one classification SVM algorithms, combining EN signals with odour concentration of gas collected on 
the arrival tank and the citizens’ reports about the presence or absence of odours nearby the plant. Results 
achieved prove that the EN is capable to reveal in real-time the deviation from normal operating conditions, and 
provide information about the potential causes of the malfunction. This allowed the activation of gas and liquid 
samplers when the anomalous condition occurred to collect the samples in a targeted way and better investigate 
the causes of the problem. This information should be useful for the identification of the substances responsible 
of odour peaks and consequently to develop a specific tailored abatement system, thus optimizing the costs 
and benefits compared to more generic solutions. 
Future goals to be achieved within this project concern the implementation of a more precise quantification 
model based on regression to estimate the odour concentration. Moreover, future works should focus on the 
validation over time of the model to spot the sensors drifting and intervene to compensate it with an appropriate 
drift correction model aimed at extending the model validity. 

References 

Abdi, H. and L. J. Williams (2010). "Principal component analysis." WIREs Computational Statistics 2(4): 433-
459. 

Collins, M. A. and L. Moy (1995). The Electronic Nose for Process Control. Neural Networks: Artificial 
Intelligence and Industrial Applications, London, Springer London. 

Leung, D. and J. A. Romagnoli (2002). Chapter 6.4 - Fault Diagnosis Methodologies for Process Operation. 
Computer Aided Chemical Engineering. B. Braunschweig and R. Gani, Elsevier. 11: 535-556. 

Majchrzak, T., W. Wojnowski, T. Dymerski, J. Gębicki and J. Namieśnik (2018). "Electronic noses in 
classification and quality control of edible oils: A review." Food Chemistry 246: 192-201. 

Martı́, M. P., O. Busto, J. Guasch and R. Boqué (2005). "Electronic noses in the quality control of alcoholic 
beverages." TrAC Trends in Analytical Chemistry 24(1): 57-66. 

Schölkopf, B., J. C. Platt, J. Shawe-Taylor, A. J. Smola and R. C. Williamson (2001). "Estimating the support of 
a high-dimensional distribution." Neural computation 13(7): 1443-1471. 

Sun, M. (2014). Support Vector Machine Models for Classification. Encyclopedia of Business Analytics and 
Optimization. J. Wang. Hershey, PA, USA, IGI Global: 2395-2409. 

 

174


	44prudenza.pdf
	Electronic Nose for Real-time Monitoring of Odour Emissions at a Wastewater Treatment Plant