Development of an electronic nose to characterize odours emitted from different stages in a wastewater treatment plant

Water Research 134 (2018) 92e100

Contents lists available at ScienceDirect

Water Research journal homepage: www.elsevier.com/locate/watres

Development of an electronic nose to characterize odours emitted from different stages in a wastewater treatment plant Andy Blanco-Rodríguez a, *, Vicente Francisco Camara b, Fernando Campo c,
n d, Vitor Debatin Vieira e, Henrique de Melo a,
n d, Alejandro Dura Liliam Bechera Alejandro Rafael Garcia-Ramirez c a Laboratory of Air Quality Control (LCQAr), Department of Sanitary and Environmental Engineering (ENS), Federal University of Santa Catarina (UFSC),
polis, SC, Brazil 88040-900, Floriano b
polis, SC, Brazil Aire Engenharia e Consultoria, 88054-340, Floriano c
gicas da Terra e Do Mar (CTTMar), Universidade Do Vale Do Itajaí (UNIVALI), 88302-202, Itajaí, SC, Brazil Centro Em Ci^ encias Tecnolo d Institute of Materials Science and Technology (IMRE), University of Havana, 10400, Havana, Cuba e
rio de Eletroforese Capilar (LabEC), Departamento de Química, Universidade Federal de Santa Catarina (UFSC), 88040-900, Floriano
polis, SC, Laborato Brazil

a r t i c l e i n f o

a b s t r a c t

Article history: Received 8 September 2017 Received in revised form 22 January 2018 Accepted 27 January 2018 Available online 3 February 2018

Wastewater treatment plants have widely been described as a significant source of odour nuisance, which has led to an increase of neighbourhood complaints. Therefore, to mitigate the negative impact of odours, the detection and analysis of these emissions are required. This paper presents a measurement system based on an electronic nose for quantitative and qualitative odour analysis of samples collected from six different stages on a wastewater plant. Hence, two features vectors were performed in order to represent quantitative trends of the gaseous mixture sampled on the facility. In addition, odour fingerprints and a PCA were computed to discriminate odours from its sources and to detect relationships among the samples. This approach also comprises a dynamic dilution olfactometer. A PLS regression model was performed to predict the odour concentration by the electronic nose in term of odour units per cubic meter. The results show that the developed electronic nose is a promising and feasible instrument to characterize odours from wastewater plants. © 2018 Elsevier Ltd. All rights reserved.

Keywords: Electronic nose Wastewater treatment plant Olfactometric analysis Environmental odours

1. Introduction Malodorous released from wastewater treatment plants (WWTPs) can cause health problems, nuisance to the community and frequently neighbourhood complaints (Carrera-Chapela et al., 2014; Gostelow et al., 2001; Stuetz and Nicolas, 2001). Indeed, odours are recently considered as atmospheric contaminants (Capelli et al., 2013). The management strategies to mitigate odour nuisance entail monitoring, assessment and controlling the generated substances. However, this issue implies significant challenges because odours are complex gaseous mixtures, which

* Corresponding author. E-mail addresses: [email protected] (A. Blanco-Rodríguez), vicente@ aireconsultoria.com.br (V.F. Camara), [email protected] (F. Campo), liliam@
n), [email protected] (A. Dura
n), [email protected] imre.uh.cu (L. Bechera (V.D. Vieira), [email protected] (H. de Melo), [email protected] (A.R. Garcia-Ramirez). https://doi.org/10.1016/j.watres.2018.01.067 0043-1354/© 2018 Elsevier Ltd. All rights reserved.

can be found at low concentrations at ambient conditions. Odours also exhibit high variability over time, which can be related to weather conditions, effluent load characteristics and others specific features. The random and temporal population activities relative to sewage disposal can also influence over odour emission (Bourgeois and Stuetz, 2002; Frechen, 2004; Jeon et al., 2009; McGinley and ~ oz et al., 2010; Stuetz et al., 1999a; Wilson, McGinley, 2008; Mun 2012). In addition to the complexity, smell perception involves a subjective interpretation that varies according to many factors. Therefore, to successfully assess odours emissions, the use of suitable measurement instruments is required. There are different techniques to analyze gases and odours. Detectors and gases analysers provide information about specific gases concentration in the odorant mixture. The more complex gas chromatography coupled to mass spectrometers (GC-MS) (Chin et al., 2017), and also with sulphur chemoluminescence detector (Sun et al., 2014) can be used to identify and quantify potential odorant compounds, usually expressed in ppm or ppb. However, it

A. Blanco-Rodríguez et al. / Water Research 134 (2018) 92e100

is not possible to characterize an odour as a whole by only analysing its individual components (Blanes-Vidal et al., 2009; Qu et al., 2008; Stuetz et al., 1999a). This kind of measurements does not encompass the gas mixture interactions (suppression, synergism, hypoadditivity and linear sum), which might lead to uncertain environmental assessment. Moreover, the gases and their respective concentrations on the odorant mixtures cause different effects on human odour perception (Capurro et al., 2012; Kim and Kim, 2014; Kuebler et al., 2011). Despite that analyses by GC-MS provide accurate and reproducible measurements, the cost of the analyses and the restriction to perform measurements in the laboratory (not in-situ) are other drawbacks that constrain the use of GC-MS. The terms odorant and odour must be explained. An odorant is a gaseous chemical component which stimulates the human olfactory system, while an odour is the sensorial response of the olfactory organ when sniffing certain volatile substances (Brattoli et al., ~ oz et al., 2010). 2011; Gostelow et al., 2001; Mun For odour measurements, the dynamic olfactometry is the more applied methodology (Capelli et al., 2008b). This technique employs an olfactometer combined with human assessors. An olfactometer is an instrument that performs controlled dilutions of a gaseous sample, then exposing them to the assessors panel and compute results. The assessors are certified experts with certain capabilities to sniff, previously selected through different tests. Then, the results of odour concentration are quantified in odour unit per cubic meter (O.U.m 3). This measuring unit represents the number of dilutions with neutral air (odourless) that are necessary to the odorous sample achieves its odour detection threshold (the concentration at which there is a 50% probability of detectability by the human assessors) (Gostelow et al., 2001). For this reason, this technique is the more appropriate method to characterize the odours released to the atmosphere. Indeed, most environmental odour regulations in different countries and municipalities are based on the odour concentration in O.U.m 3 (Brancher et al., 2017). Although, olfactometry is expensive, time-consuming, and presents lower repeatability and accuracy due to its subjective nature (Brattoli et al., 2011). The so-called electronic nose (e-nose) is another kind of instrument, which can be employed for odours assessment (Capelli ~ oz et al., 2010). The e-nose, in a simple way, et al., 2008b; Mun mimics the mammalian olfactory system in term of sensory response and information processing (Arshak et al., 2004). These instruments mainly contain an array of sensors with crosssensitivities, and an appropriate patterns recognition system capable of recognising simple or complex odours (Gardner and Bartlett, 1994). There are some relevant features that distinguish the e-noses regarding the rest of the gas/odour measurement systems: these concomitantly support gases and odours analyses, as well as temperature, humidity, wind velocity and others variables (Abdullah et al., 2012; Dentoni et al., 2012); allowing continuous monitoring of input odour data; and also performing both qualitative and quantitative analysis. For continuous odour monitoring can be used to estimate odour impact on the neighbourhood in real time, support rapid information for population, acquire data of high odour peaks over short time scales, capture of extreme odour events and even for a proactive purpose to detect odours before their impact on surrounding areas (Bourgeois and Stuetz, 2002; Capelli et al., 2008a; Purenne et al., 2007). A relate difficult from e-noses is that the non-specific gas sensors can respond to both odorous and odourless substances. In fact, this is a drawback of enoses, not only for environmental odour applications. The low sensitivity of gas sensors to the odours threshold is another problem of e-noses (Boeker, 2014). However, to environmental odour analysis, these systems are an interesting choice, and they have

93

been successfully applied in several assessments. The development of an electronic nose comprises different stages that embrace the selection of the sensor array and the conditioning circuits, the processing and signal acquisition hardware as well as the signal processing, training and analysis of data. As concerning the selection of sensors, each sensor should maximize the overall sensitivity, providing different selectivity profiles over the range of application to the target odour (Phaisangittisagul et al., 2010). Then, a first step consists in obtaining several features from each sensor dynamic response, validating them and picking the main features which characterize the odours under study. This selection of the sensors strongly depends on the application, and it is needed when classification performance, cost, and technology limitations are issues of concern (Phaisangittisagul et al., 2010). The next stages in processing the e-nose data include: data signal-pre-processing, feature extraction, feature selection, classification, regression, clustering, and validation. That way, several methods from statistical pattern recognition, artificial neural networks, chemometrics, and machine learning has been used to process electronic nose data (Gutierrez-Osuna, 2002). Frequently, gas/odour monitoring systems only comprise one kind of measurement device. Nevertheless, it is also relevant to correlate outputs from different instruments. This integration usually provides more detailed and encompassed outcomes ~ oz et al., 2010; Sohn (Abdullah et al., 2012; Brattoli et al., 2011; Mun et al., 2008). The analysis of environmental odours demands some difficult tasks. Hence, it can be supported by the combination of results from various measurement instruments, in order to obtain more representative data about substances evaluated. For instance, coupling olfactometry with GC-MS (GC-MS-O) allows the identification of odour-active compounds, which indicate the relevance of some gases as odorants. For this demand emerge another possibility, correlations between e-noses and olfactometry, which can allowed portable and fast odour analysis in term of odour concentrations. It can be used as a public tool to attend cases of complaints or to evaluate odour episodes that cause impacts on the populations (Brattoli et al., 2011; Purenne et al., 2007). Several measurement instruments have been used for environmental applications (Alam and Saeed, 2013; Bootsmaa et al., 2014; Capelli et al., 2013; De Melo Lisboa et al., 2009; Wilson, 2012), including the assessment of odours emitted from WWTPs. There are reported e-noses responses correlated with olfactometry analysis to supply quantitative results from wastewater odours (Purenne et al., 2007; Zarra et al., 2014). Guz et al. (2015) performed a comparison of e-nose response to the standard physical-chemical parameters of treated wastewater, while Zarra et al. (2009, 2014) compared and evaluated different odour measurement methods for wastewater odours. Rajbansi et al. (2014) presented an assessment of odours from a sewage wastewater in terms of odour intensity by human assessors and GC-MS. A portable and commercial e-nose (PEN3) in conjunction with GC-MS was employed to discriminate between alkaline-stabilized biosolids treated at different doses (Romero-Flores et al., 2017). A sophisticated network of e-noses to quantify odours at Montreal WWTP was reported by Purenne et al. (2007). This system used correlations with olfactometry to calibrate the e-noses and also employed these results as input data to an atmospheric dispersion model. Various e-noses have also been applied as a single instrument to wastewater odours assessment. Capelli et al. (2008a) proposed a system with three e-noses for continuous monitoring of environmental odours at specific receptors around a WWTP. In this approach was identified the major odour source in the facility and was also estimated the odour impact on the neighbourhood. The enose presented by Stuetz et al. (1999b) was able to discriminate odour samples from three different sources: raw sewage, settled

94

A. Blanco-Rodríguez et al. / Water Research 134 (2018) 92e100

sewage, and final effluent. Franke et al. (2009) represented qualitative differences among sewer samples through sensors profiles € features in a polar graph. While Oztürk et al. (2009) obtained an intensity response of quartz crystal microbalance (QCM) sensor array. Then, the peaks of these input signals were used to infer a primary odours discrimination among samples from biological treatment unit, sedimentation tank and inlet unit. Nicolas et al. (2000) implemented a classification model through Discriminant Analysis (DA), and Bourgeois and Stuetz (2002) used PCA to cluster samples collected on the inlet of a WWTP. Another application was presented by Nake et al. (2005), which employed two portable commercial e-noses to analyze samples of outdoor locations near a wastewater facility. However, the three data classes identified were overlapping and also the observations were spread in the PCA graph, making it difficult to establish appropriate clusters. For qualitative assessment of environmental odours, the e-noses can use various processing tools. PCA is one of commonly method used to discriminate odour patterns. Through the PCA the input data with n-dimensions is transformed in a new dimensional space, where each dimension (axes) is orthogonal to the others. The new axes, called principal components, are linear combinations of input data and these are uncorrelated. Thus, the principal components are organized decreasingly, from maximum to the minimum variance, that can be adequately computed. Graphically, the first axis is located in the direction of greatest data variation, and the second perpendicularly. Then, the variables under study can be represented in the PCA space, plotted by the so-called scores in a 2D or 3D graph. Afterward, data patterns can be established according to the proximity of scores (Bro and Smilde, 2014; Haykin, 2005; Romero-Flores et al., 2017). Linear Discriminant Analysis (LDA) is another tool that can be employed to distinguish data groups. Basically, this supervised classification method allows achieving an optimal classes separation from a linear combination of data. This procedure is based on creating linear functions to maximizing the ratio of the between-group variance to the within-group variance. This criterion is the main difference with PCA. Therefore, from the linear function, target data can be classified in term of classes (Boeker, 2014; Romero-Flores et al., 2017; Sironi et al., 2007). Despite the advantages and potential of the e-noses, the use of these instruments to carry out an extensive analysis of wastewater odours has not been completely explored. For instance, to discriminate the odours sources inside a WWTP in a comprehensive way. This kind of assessment could be useful to identify sources responsible for odour nuisance events. Consequently, a localized control action could be executed, in order to restore the correct operation of the process affected or even to adjust some plant functionalities for proactive purposes. This study presents the application of an e-nose, for qualitative and quantitative odour analysis from a WWTP. Odour samples from six different stages in a wastewater facility were analysed. The data acquired was processed by the e-nose in order to establish

quantitative vectors features along the treatment stages and to discriminate the samples’ origin. A dynamic dilution olfactometer combined with human assessors was also used. A correlation between the e-nose and the olfactometer results was modelled by PLS to predict the odour concentration. 2. Materials and methods 2.1. Electronic nose analysis The e-nose is based on a previous system developed in the laboratory (Valiente Romero, 2014). The instrument comprises a gas sensor array, an automatic fluid transportation line, a microcontroller board and software processing tools to carry out qualitative and quantitative analysis of odour samples. To select the gas sensors, the following criteria were considered: high sensitivity to the main odorants emitted from WWTPs; crosssensitivity sensors (suitable for e-nose analyses); minimal number of sensors and measurement channels for less complexity and cost of the system, and because excessive number of sensors can increase noise and processing time without useful information; previous studies of similar sensor array configuration for this kind of application which have been tested with successful results; and use of commercially available sensors. Then, in this approach, five MOS sensors manufactured by Figaro (Figaro Inc., Japan) were used: TGS2611, TGS2602, TGS2610, TGS826 and TGS2600. Resistive MOS sensors are relatively small size, low cost and these have been widely used to detect odorants from sewage treatment activities (Capelli et al., 2008a; Guz et al., 2015; Nicolas et al., 2000; Purenne et al., 2007; Zarra et al., 2014). As recommended by the manufacturers, these type of sensors requires a simple conditioning circuit. Hence, to measure the sensor resistance output a typical resistive voltage divider was used. The input gases induce chemical changes on the sensing surface, which causes variations in the electrical conductivity of these devices. According to the manufacturer (Figaro, 2004a; b; 2005a; b; 2006), in Table 1 are presented the sensors employed and their corresponding sensitivities to the gases. This sensors' response capability matched with some of the major odorants emitted from WWTPs, mainly H2S and NH3. In addition, by the sensors used on the e-nose, the emission of other odorants can be indirectly measured by chemicals that are byproducts of anaerobical activity (such as methane, ethanol and hydrogen) (De Mes et al., 2003). MOS sensors also present crosssensitivity or non-selectivity, which indicates that the sensors respond towards different input compounds. The voltage values from the five sensors were sampled with an inbuilt analog to digital converter (ADC) in the Arduino Mega 2560 microcontroller board (ATMEL, U.S.A.). This board is coupled, via RS232, with a Virtual Instrument (VI) developed in LabVIEW platform (National Instruments, USA). This VI running on the computer

Table 1 Gas sensitivities of the MOS sensor array used in the e-nose. Sensor name

Sensor sensitivity

TGS2611 (Methane gas) TGS2602 (Air contaminants) TGS2610 (LP gas) TGS826 (Ammonia gas) TGS2600 (Air Quality)

Methane; Iso-butane; Hydrogen; Ethanol Toluene; Hydrogen sulphide; Ethanol; Ammonia; Hydrogen Iso-butane/Propane; Methane; Hydrogen; Ethanol Ammonia; Ethanol; Hydrogen; Iso-butane Ethanol; Hydrogen; Iso-butane; Carbon monoxide; Methane

A. Blanco-Rodríguez et al. / Water Research 134 (2018) 92e100

to command the acquisition process, stores data from sensors, shows graphical information about electrical profiles and manages the fluid transportation line. The fluid transportation line is mainly composed by three solenoid valves; teflon and silicone tubing; a sensors chamber; a charcoal filter to obtain the baseline from ambient air; and a vacuum pump. This delivery block allows establishing a four-stage sequence for measurement: baseline (passing reference gas), rise transient (passing odour), steady state (maintaining odour inside the sensor chamber) and recovery transient (passing reference gas again). The amplitude versus time response of a gas sensor has a pulse shape output, which is obtained by modulating the passage of the sample and the odourless reference through the sensors chamber. This dynamic response from an array of sensors contains qualitative and quantitative information to characterize the gas under analysis. A typical modulation profile comprises the four steps mentioned above: baseline, rise transient, steady state and recovery transient (Arshak et al., 2004). A sequence of operations to deliver the gas under analysis, from the odour sample to the sensors chamber, is carried out. Data acquisition from sensors output is performed from the beginning of this sequence. Duration of these steps can vary from seconds to minutes in the e-noses based on MOS sensors (Arshak et al., 2004). The time duration of steps must be sufficient for the sensors to reach the steady stage for all odour sources and to return to the baseline. In addition, consuming the minimum volume of the sample is required because olfactometry analysis demands tens of liters. 2.1.1. Data pre-processing After pulses acquisition, a data pre-processing stage was carried out. This is a basic stage in the e-noses because the sensors outputs can present drifts and redundant information, which should be treated before performing the instrument response. By this procedure, the signals from sensor array, VRL, were adjusted and main features were extracted. Data pre-processing in e-noses mainly involves signal filtering, baseline manipulation, normalization and feature extraction (Campo García, 2016; Ferreira, 2015; Marco and
rrez-Ga
lvez, 2012). Gutie For this application, a moving average filter was applied to each voltage data profile in order to remove random errors (Smith, 1999). Then, a baseline manipulation differential method was computed by the difference between each digital sample of the profile and the baseline value, thus reducing noise and sensor drifts (Arshak et al., 2004; Gutierrez-Osuna and Nagle, 1999; Pearce et al., 2003). Afterward, the Euclidian norm for data fitting to nondimensional values was determined. It allows correcting the scale difference between dataset, what is a requirement for any subsequent data comparison (Pearce et al., 2003). During feature extraction, different parameters from pulses with higher discrimination capability were computed and picked, for example the maximum voltage (Vmax) and rising transient slope. To perform this data pre-processing stage, a software code was developed in Matlab (2017a) (MathWorks, U.S.A.) as specific applications tool. 2.1.2. Odours analysis Quantitative and qualitative analyses were performed to characterize gaseous mixture and odour emissions from a WWTP. Firstly, a quantitative trend of the sensors responses across the treatment stages in the facility was associated with the compounds released. For this purpose, two features vectors, called V1 and V2 were computed from the Vmax of the electrical profiles. Both vectors were defined with six elements each, performing one element from treatment stage. Then, V1 was a vector of means, with each element computed from Vmax means of all sensors, at each location. While

95

the V2 was another vector of means, performed from Vmax means of the three sensors with the highest responses, at each location too. By this procedure, it is possible to represent, in a simple way, the emission of the gaseous mixture as a whole, due to the nonselectivity response of the MOS sensors. Then, from V1 and V2 performed at each treatment stage, relationships about the gaseous mixtures emissions among different locations of the plant were obtained. A qualitative analysis by odour fingerprints and PCA was computed to discriminate odours from its source in the WWTP. PCA is an unsupervised multivariate technique, which describes the data in term of variance. The scores produced, represent the maximum variance, and it may be plotted in few dimensions. This tool is mainly used to identify data clusters as well as to reduce the dimensionality by a linear compression of data. PCA is a simple and fast method, and it is useful to perform a qualitative analysis of
rrez-Ga
lvez, 2012; Scott et al., 2007). odours (Marco and Gutie Regarding quantitative analysis, a correlation between e-nose and olfactometric outputs was also computed to estimate MOS sensors response in O.U.m 3. This relationship was modelled by a PLS regression technique, which is frequently used in the chemical
rrez-Ga
lvez, 2012; Wold et al., 2001). sensor array (Marco and Gutie Synthetically, PLS is a linear multivariate method for relating two groups of data by regression coefficients. Each group, expressed as a matrix, represents independent and dependent variables. In this case, the feature extracted from electrical profiles at the preprocessing stage is considered the independent variable while the olfactometric test output is the dependent variable. So, the regression coefficients were firstly computed from e-nose and olfactometry responses. Typically, these data set are called as training or calibration set. Afterward, it is possible to predict new input data in term of O.U.m 3 by the training matrix modelled (Wold et al., 2001). The correlation coefficient R-square and the root mean square error (RMSE) were used to evaluate the performance of the predictive model. The coefficient R-square indicates the existence of a linear relationship between model variables, while RMSE measures how well the model adjusts the data. 2.2. Olfactometry analysis The olfactometric measurements were performed by using a dynamic dilution olfactometer Odile 3500 (Odotech, Canada), according to the European Standard EN 13725 (CEN, 2003). The instrument is composed by an air purification system, a pressurized vessel, a dilution unit, a six-cell olfactometric table and an operation software running on a personal computer. Dilutions were made decreasingly and logarithmically, and then exposed to a panel of six certified experts. The gas concentrations were quantified in odour units per cubic meter, O.U.m 3. 2.3. Experimental procedure The study was carried out in a three-stage procedure: odour sampling, measuring and data processing. In the first stage, odour samples from a wastewater facility were collected in sampling bags and carried to the laboratory. At the second stage, each sample was measured by the electronic nose and by the olfactometer. According to EN 13725 standard (CEN, 2003), a 30 h time lapse between sampling and measuring stages was respected. Finally, acquired data were processed in order to obtain the response of the instrument. 2.3.1. Odours sampling The odour sampling was conducted on a WWTP in Criciúma, southern Santa Catarina State, in Brazil. The WWTP is composed by

96

A. Blanco-Rodríguez et al. / Water Research 134 (2018) 92e100

a preliminary system with screening and grit chamber, flotation tank, two parallel Upflow Anaerobic Sludge Blanket (UASB) reactors, and three parallel settling tanks. Firstly, were identified six locations as potentially the main sources of odours within the WWTP. These sampling points were located on the exhaust of the preliminary treatment, which leads to biofilter inlet (BioInlet); after the biofilter (BioOutlet); on the flotation tank (Flotation); on the exhaust line of one UASB reactor (Flare); on a settling tank (Settling); and on a sludge thickener (Sludge). Sampling was carried out according to the source type. BioInlet, BioOutlet and Flare were point sources (where direct sampling was used). Flotation, Settling and Sludge were area sources (flux chamber sampling was applied). A preliminary sampling test was performed on 27/10/2016. For this purpose, only three sampling bags on the area sources were collected (Flotation, Settling and Sludge). Then, two sampling campaigns were conducted under different WWTP operating regimes to more broadly characterize the emissions. Odour sampling was carried out on 31/10/2016 (Campaign 1) and 10/11/2016 (Campaign 2). The Campaign 1 was conducted on a Monday, which according to the WWTP operators is a critical day, due to the habit of the population of washing cars and clothes on the weekends. In fact, a lot of foam was observed during sampling on that day. The household detergents (sulphonates) influence in the organic sulphur content of sewage effluents (Gostelow et al., 2001). While the Campaign 2 occurred on a Thursday, which is considered as a normal operating day. One sampling bag was collected on each point for each campaign. Different equipment, materials and instruments were selected to odour sampling procedure. Vacuum pumps, model MOA (Gast Manufacturing, USA), were employed. These have a metallic interior, are oilless and were cleaned prior to sampling. Polyurethane and silicone tubing, which have inert and odourless features were used, as well as glass and teflon valves coupled to sampling bags. Nalophan® bags, with a capacity of 60 L approximately, were employed. This kind of bag preserves the sample gas composition during a certain time due to its adsorption properties. In addition, at the three area sources, the flux chamber method was applied. This device was conceived as stated on EPA's document (Klenbusch, 1986). Two flowmeters, model VFA-24 (Dwyer Visi-Float®, USA), two thermocouples (type K) and an adsorption system (activated charcoal, to clean the sweep air) were also used.

2.3.2. Odours measuring Each sampling bag collected at the previous stage was measured

in a sequence by the e-nose and olfactometer, described above. Then, five replicates per bag were selected and five electrical profiles per sensor were performed via the e-nose, obtaining 25 pulses per odour sample. The time intervals for each profile sequence were experimentally determined as 50 s (baseline), 50 s (rise transient), 1 s (steady state), and 75 s (recovery transient). This measuring stage was performed in the Laboratory of Air Quality Control (LCQAr) at Federal University of Santa Catarina (UFSC).

3. Results and discussion 3.1. Sensor array response by the e-nose In Fig. 1, the sensors’ response to odours from six sources is synthetically presented. The bars amplitude represents the maximum voltage (Vmax) in term of means and standard deviations, computed from each odour sample in Campaign 1. The Vmax parameter was the feature extracted at the pre-processing stage. The names of locations displayed in Fig. 1 were ordered on the Xaxis according to the sequence of treatment stages in the WWTP, starting from BioInlet to Sludge. In the previous graph, each sensor showed readable output signals, expressed by the amplitudes of the Vmax means, registered between 0.05 V (TGS2600-Sludge) and 2.18 V (TGS2610-Flare). From the previous figure can also be observed low dispersion values around the means, represented by standard deviations less than 0.06 V. On that point, only the responses from TGS2611, TGS2610 and TGS2600 to the Flare samples, showed highs values of standard deviation, equals to 0.41, 0.41 and 0.28 V, respectively. It was caused by one profile (for each sensor), which presented slightly higher amplitude than the other four profiles. In addition, a cross-sensitivity response of the sensors to the target odours can be confirmed, which is indicated by the broad variability among the Vmax sensors outputs. It suggests that the instrument is capable to acquire wide and representative input data from the samples collected. This behaviour is verified through the large amplitude difference among: the same sensor across all treatment stages, and all sensors within each stage. Being that, only two minimal differences were computed. These values were 0.01 and 0.001 V for (TGS2602-Flare)-(TGS2602-Settling) and (TGS2610-Sludge)(TGS826-Sludge), respectively. Also, the Vmax value of each sensor to the BioInlet was higher than the BioOutlet, which implies on gas concentrations reductions. It suggests a proper biofilter treatment in terms of gaseous compounds removal.

Fig. 1. Response of the sensor array to odours released from WWTP in Campaign 1.

A. Blanco-Rodríguez et al. / Water Research 134 (2018) 92e100

97

Fig. 2. a (left) and 2b (right). Features by the e-nose response across WWTP samples: a) in Campaign 1; b) in Campaign 2.

Different profiles from sensor responses were found among sampling locations. For example, TGS2602 sensor showed different amplitude levels between the exhaust line of the preliminary treatment (BioInlet and BioOutlet) and the other odour sources. It suggests a greater presence of hydrogen sulphide and/or ammonia in comparison to the other sampling sites. From Fig. 1 can also be observed that higher sensors' reactions were obtained in Flotation and Flare points. Sensors that stood out were TGS2610, TGS2611 and TGS2600. This response might be linked to the turbulence caused on the Flotation tank (which leads to gases release from the liquid phase). Higher reactions on the Flare point might be due to its nature, it exhausts gases from a UASB reactor, which induces microbial anaerobic activity, and the generation of gases/vapours such as methane, hydrogen and ethanol. To quantitatively represent the concentration of the gaseous mixture across the treatment stages, the feature vectors V1 and V2 were presented in Fig. 2a and 2b. From Fig. 2a and b, the V1 and V2 showed a decrease between the beginning (BioInlet) and the end (Sludge) of the process in the WWTP, mainly in the Campaign 1. In general, Campaign 1 responses were higher than responses of Campaign 2. The Flare point with the higher response it might be related to the fact that Flare supply a gaseous outlet from a UASB reactor. Since these anaerobic digesters have a tendency to emit odorants and gases. Therefore, it is expected the most significant sensor responses for on Flare source. The smallest values were observed in the last stages of the treatment, which makes sense because most of the organic matter (that could possibly be digested by bacteria and emit odours) was removed by the previous reactor (UASB). Similar profiles can observe by comparing V1 and V2 in each graph, and V1 vector presented lower amplitude than V2 for all treatment stages. Thus, V1 feature vector provides a moderate response along the sampling points and it can be useful, for example, to represent the operation regime in the facility. While V2 presents more sensibility, which may aid to report possible events of sudden increases from gases. Various parameters, performed from V1 and V2 features were

presented in Table 2. The relationship between Vmax from Sludge and Vmax from BioInlet was presented. In Campaign 1, it was 10.1 and 10.2% for V1 and V2, respectively. While in Campaign 2, the relationship was 35.7 and 45.1% for V1 and V2, respectively. Then, from this parameter was also estimated the gaseous mixture reduction by the plant. In Table 2 the treatment of the gaseous mixture by the biofilter was also presented. It was estimated through the relationship between Vmax from BioOutlet and Vmax from BioInlet. These results were close for both campaigns, ranged between 85.8 and 87.7%. Then, the gases reduction by the biofilter was also estimated, indicating that its performance was stable, around 13.4%. This parameter also suggests a gaseous mixture removal by this treatment stage of approximately 13% for both campaigns. These results do not provide information in terms of odour concentrations, for this purpose a multivariate regression was developed subsequently. 3.2. Odour fingerprints The maximum of the electrical profiles from each sensor, Vmax, was represented in two polar graphs, Fig. 3a (Campaign 1) and Fig. 3b (Campaign 2). The Vmax value showed in both figures was the compute average from five replicates. This magnitude was plotted in five axes, one per sensor. Then, as each sensor is sensitive to various gaseous compounds, the shapes obtained are visual qualitative indicators of the samples, called fingerprint. From the previous figures, can be perceived different shapes, which represent odour samples points according to their original location in the plant. Indeed, a clear pattern variation among the samples were displayed. In Campaign 1, Fig. 3a, is noteworthy the differences among the shapes. Only BioInlet and BioOutlet treatment stages were similarly represented. In Fig. 3b (Campaign 2) different patterns can also be observed, mainly from Flare source. However, the radius of Vmax decreased and some shapes were alike. For instance, Settling-Sludge and Flotation-BioOutlet. Visualizing Vmax data through a relative sensor response allows

Table 2 Parameters computed from V1 and V2 features. Campaign 1

Higher Vmax (V) Vmax (Sludge)/Vmax (BioInlet) (%) 100 - Vmax (Sludge)/Vmax (BioInlet) (%) Vmax (BioOutlet)/Vmax (BioInlet) (%) 100 - Vmax (BioOutlet)/Vmax (BioInlet) (%)

Campaign 2

Feature V1

Feature V2

Feature V1

Feature V2

1.41 10.1 89.9 85.8 14.2

1.96 10.2 89.8 87.7 12.3

1.10 35.7 64.3 86.5 13.5

1.47 45.1 54.9 86.6 13.4

98

A. Blanco-Rodríguez et al. / Water Research 134 (2018) 92e100

Fig. 3. a (left) and 3b (right). Polar graphs of the relative sensors responses: a) computed from samples of Campaign 1; b) computed from samples of Campaign 2.

to distinguish among different odour sources from a WWTP. Nevertheless, to more explore the sensor array response another qualitative odour analysis by a PCA was developed. 3.3. Discrimination of WWTP odours The Vmax was also the parameter used as input data to the PCA model. Results of these unsupervised technique were presented in Fig. 4a and Fig. 4b. In these graphs, each plot represents its corresponding sensor signal, being that five pulses per sample were acquired. Through Fig. 4a is possible to discriminate six clusters, related to the six odours sources sampled in the Campaign 1. The plots of each group were close and all groups were located separately, which led to identifying six odour patterns. From Fig. 4b, can also distinguish the same number of clusters. However, the observations of Settling and Sludge were close, almost overlapping. A hypothesis about this proximity is a difference between the effluents features and/or WWTP operating regime at Campaign 2 regarding Campaign 1. Sampling weather conditions in term of temperature and relative humidity were similar for both campaigns. Therefore, these variables must not have influenced in a significant way for this odour assessment. Also, for Campaign 2 the plots from Flotation samples were spread in the graph, but even so, they were disjointed from the others observations. 3.4. Prediction of odour concentration To estimate the odour concentrations from WWTP emissions a

prediction model by PLS regression technique was applied. To achieve more broad input data, in this analysis was also include the responses from the samplings bags of the preliminary test. Averages of Vmax and rising transient slope from the five replicates of each sensor were the parameters selected as one of the input variables to the PLS model. The other input was the odour concentration value obtained by the olfactometric assessments. Related to the e-nose, a data input matrix with 15  10 (observations x variables) elements was arranged. The columns contain orderly the five Vmax values and the five transient slopes values for each sensor. While in the rows were fixed these variables for the 15 sampling points. The olfactometry input data were organized in a 15  1 matrix, where each row represents their corresponding observation in O.U.m 3. From these data, the regression coefficients were computed and then the predicted model output was computed as odours concentrations values, Fig. 5a and Fig. 5b. In Fig. 5, blue circles represent the modelled odours data, the blue dashed line shows the fit and the black dashed lines represent the error bounds for a 95% of confidence level. From Fig. 5a was considered three observations as outliers (black circles) due to these points were out of scale. Then, by excluding the anomalous values the fit was improved, Fig. 5b. The predicted concentrations are distributed around the ideal line 1/1. It was confirmed by slope and intercept values with the Y-axis of fitting line, equals to 0.9597, and 410.8 O.U.m 3, respectively. However, a slight skewness around the fitting line was perceived. The value of R-square was 0.9967 for a 95% of confidence level, which showed a linear dependence. The coefficient RMSE was equal to 1.17  104, indicating that from the model output is

Fig. 4. a (left) and 4b (right). PCA data analysis to identify odours patterns regarding its locations inside the WWTP: a) PCA computed from samples of Campaign 1; b) PCA computed from samples of Campaign 2.

A. Blanco-Rodríguez et al. / Water Research 134 (2018) 92e100

99

Fig. 5. a (left) and 5b (right). Predicted odour concentration performed by a PLS model: a) all dataset with three outliers represented by black circles; b) excluded the three outliers.

expected an error range around this value. Due to the RMSE value represents only the 0.17% of the measurements range, it can be considered appropriate for this application. Through the PLS components used in the multivariable model, most of the variances in predictor variables and response matrix was explained. The PLS model performed was able to establish a linear relationship between the e-nose output and the olfactometry analysis. Then, a global calibration curve for the whole plant was obtained. It allows to predict the odour concentrations, released from all treatment stages in the WWTP, by the e-nose. 4. Conclusions In this study, a WWTP odours samples characterization by using an e-nose was presented. Samples collected from six locations, which represented the main odour emitting sources in the plant, were measured: BioInlet, BioOutlet, Flotation, Flare, Settling and Sludge. Both, quantitative and qualitative analysis were performed. By this approach, some steps for analysing the samples from WWTP were proposed: - Setting of trend functions to represent the emitted gaseous mixture across the plant. For this purpose, two features vectors, called V1 and V2 were proposed. Both features represent a quantitative approximation about compounds released, in term of Vmax variations. These vectors can be useful to represent the operation regime of the plant and to report possible events of sudden increases from substances. Through this multi-sensor analysis response is possible to establish a primary assessment about the gaseous mixture in a simple way, which can assist to estimate and compare the compounds emitted from specific locations in the WWTP. - Calculation of parameters from V1 and V2 to estimate the efficiency in the treatment of the gaseous mixture by the plant and through the biofilter. Then, a relationship between Sludge (final treatment stage) and BioInlet (first treatment stage) of emitted compounds was presented. Others relationship could be computed, like at biofilter stage. In addition, the reduction of gaseous substances by the plant was estimated, which suggests a removal of approximately 90 and 60% of compounds emitted, in Campaign 1 and in Campaign 2, respectively. At the biofilter treatment stage, the gases reduction was around 13% for both campaigns. - To perform odour fingerprints and PCA qualitative analyses of compounds emitted at each stage. These tools allowed to identify odours patterns according to the treatment stage in the

facility. Basically, these assessments would help to identify: possible fails at WWTP locations, odours emissions, emergence of hazardous substances or other anomalies in the wastewater treatment process. - Prediction of odour concentrations in term of O.U.m 3 to provide a quantitative response according to the European Standard EN 13725. For this purpose, a PLS regression between the e-nose and the olfactometry analysis was performed. The model output showed an appropriate performance with R-square and RMSE equals to 0.9967 and 1.17  104, respectively. The results showed that the e-nose proposed is capable of measuring odour samples from a WWTP with no large input data. As concerning as the results from feature vectors, odour fingerprint and PCA, this instrument can be employed as a single device for qualitative and quantitative analysis. The instrument can provide some parameters to monitor the efficiency of odour treatment stages (as seen on biofilter). As a result, the e-nose could be a suitable tool to perform early warning responses related to odours released in the wastewater plant. This instrument allows to infer and register irregular operations in the stages of wastewater treatment and consequently take control actions, even, in a proactive way to decrease the unpleasant odours released. By the enose, anaerobic degradation in the stages of the facility can also be indirectly identified. It would assist the odours management of the plant, in order to establish proper maintenance cycles and even to adjust the dimensions of each treatment stage. This approach supports a broad response as a combination of results in term of odour concentration and qualitative information. It would be a fast and low-cost tool, projected to recognize irregularity events. Emphasizing that, for this early warning purpose, it is not necessary the use of precise and complex instruments. Additionally, the enose could be used as an in-situ instrument to monitor the odours from a WWTP. Further samples analysis at different seasons and under different WWTP operating conditions could be performed. More physico-chemical information about the sewage effluents would be handled in order to associate it with the gaseous compound emissions. Acknowledgments The authors thank Project CAPES-MES 139/11 “Desenvolvi^nicos para a detecça ~o de substa ^ncias mento de narizes eletro gasosas no meio ambiente: contribuiç~ ao para a avaliaç~ ao do ^nio de Po
simpacto de odorantes”, Programa de Estudantes-Conve

100

A. Blanco-Rodríguez et al. / Water Research 134 (2018) 92e100

~o (PEC-PG) de CAPES/CNPqeBrasil and Apoio ao PesquiGraduaça sador Visitante - APV, Proceso: 451000/2017-0, de CNPq. References Abdullah, A., Shakaff, A., Adom, A., Zakaria, A., Saad, F., Kamarudin, L., 2012. Chicken farm malodour monitoring using portable electronic nose system. Chem. Eng. Trans. 30, 55e60. Alam, H., Saeed, S.H., 2013. Modern applications of electronic nose: a review. Int. J. Electr. Comput. Eng. 3 (1), 52e63. Arshak, K., Moore, E., Lyons, G.M., Harris, J., Clifford, S., 2004. A review of gas sensors employed in electronic nose applications. Sens. Rev. 24 (2), 181e198. Blanes-Vidal, V., Hansen, M.N., Adamsen, A.P.S., Feilberg, A., Petersen, S.O., Jensen, B.B., 2009. Characterization of odor released during handling of swine slurry: Part I. Relationship between odorants and perceived odor concentrations. Atmos. Environ. 43 (18), 2997e3005. Boeker, P., 2014. On ‘electronic nose’methodology. Sensor. Actuator. B Chem. 204, 2e17. Bootsmaa, S., Leuwerinkb, T., Bilsenc, I., 2014. Online air monitoring with enoses at the tata steel plant in the Netherlands. Chem. Eng. Trans. 40, 79e84. Bourgeois, W., Stuetz, R., 2002. Use of a chemical sensor array for detecting pollutants in domestic wastewater. Water Res. 36 (18), 4505e4512. Brancher, M., Griffiths, K.D., Franco, D., de Melo Lisboa, H., 2017. A review of odour impact criteria in selected countries around the world. Chemosphere 168 (Suppl. C), 1531e1570. Brattoli, M., de Gennaro, G., de Pinto, V., Loiotile, A.D., Lovascio, S., Penza, M., 2011. Odour detection methods: olfactometry and chemical sensors. Sensors 11 (5), 5290e5322. Bro, R., Smilde, A.K., 2014. Principal component analysis. Anal. Math. 6 (9), 2812e2831. ~o de te
cnicas de pre
-processamento para um nariz Campo García, F., 2016. Avaliaça ^ nico baseado em sensores MOS. Dissertaça ~o de Mestrado. Universidade eletro do Vale do Itajaí, Itajaí, p. 105.
ntola, P., Del Rosso, R., Grande, M.I., 2008a. Electronic noses Capelli, L., Sironi, S., Ce for the continuous monitoring of odours from a wastewater treatment plant at specific receptors: focus on training methods. Sensor. Actuator. B Chem. 131 (1), 53e62.
ntola, P., Il Grande, M., 2008b. A comparative Capelli, L., Sironi, S., Del Rosso, R., Ce and critical evaluation of odour assessment methods on a landfill site. Atmos. Environ. 42, 7050e7058. Capelli, L., Sironi, S., Del Rosso, R., Guillot, J.-M., 2013. Measuring odours in the environment vs. dispersion modelling: a review. Atmos. Environ. 79, 731e743. Capurro, A., Baroni, F., Olsson, S.B., Kuebler, L.S., Karout, S., Hansson, B.S., Pearce, T.C., 2012. Non-linear blend coding in the moth antennal lobe emerges from random glomerular networks. Front. Neuroeng. 5, 6. Carrera-Chapela, F., Donoso-Bravo, A., Souto, J.A., Ruiz-Filippi, G., 2014. Modeling the odor generation in WWTP: an integrated approach review. Water, Air, Soil Pollut. 225 (6), 1e15. CEN, European Committee for Standardization, 2003. Air Quality - Determination of Odour Concentration by Dynamic Olfactometry, EN 13725. British Standard, Brussels. Chin, S.-T., Eyres, G.T., Marriott, P.J., 2017. Springer Handbook of Odor. Springer, pp. 47e48.
, T., Guy, C., 2009. Aplicaço ~es do nariz eletro ^nico nas De Melo Lisboa, H., Page
gicos em Engenharia 5 (2), indústrias e na gest~ ao de odores. Estudos Tecnolo 195e211. De Mes, T., Stams, A., Reith, J., Zeeman, G., 2003. Methane production by anaerobic digestion of wastewater and solid wastes. Bio-methane & Bio-hydrogen 58e102. Dentoni, L., Capelli, L., Sironi, S., Del Rosso, R., Zanetti, S., Della Torre, M., 2012. Development of an electronic nose for environmental odour monitoring. Sensors 12 (11), 14363e14381. Ferreira, M.M.C., 2015. Quimiometria. Franke, W., Frechen, F.-B., Giebel, S., 2009. H2S, VOC, TOC, electronic noses and odour concentration: use and comparison of different parameters for emission measurement on air treatment systems. Water Sci. Technol. 59 (9), 1721e1726. Frechen, F.-B., 2004. Odour emission inventory of German wastewater treatment plants-odour flow rates and odour emission capacity. Water Sci. Technol. 50 (4), 139e146. Gardner, J., Bartlett, P., 1994. A brief history of electronic noses. Sensor. Actuator. B Chem. (18e19), 211e220. Gostelow, P., Parsons, S.A., Stuetz, R.M., 2001. Odour measurements for sewage treatment works. Water Res. 35 (3), 579e597. Gutierrez-Osuna, R., 2002. Pattern analysis for machine olfaction: a review. Sens. J. IEEE 2 (3), 189e202. Gutierrez-Osuna, R., Nagle, H.T., 1999. A method for evaluating data-preprocessing techniques for odour classification with an array of gas sensors. Sys. Man, Cyber. Part B: Cyber. IEEE Trans. 29 (5), 626e632.
d, G., Jaromin-Glen
, K., Suchorab, Z., Sobczuk, H., Bieganowski, A., 2015. Guz, Ł., Łago

Application of gas sensor arrays in assessment of wastewater purification effects. Sensors 15 (1), 1e21. Haykin, S., 2005. Neural Networks. A Comprehensive Foundation. Hamilton, Ontario, Canada, pp. 414e464. Jeon, E.-C., Son, H.-K., Sa, J.-H., 2009. Emission characteristics and factors of selected odorous compounds at a wastewater treatment plant. Sensors 9 (1), 311e326. Kim, K.-H., Kim, Y.-H., 2014. Composition of key offensive odorants released from fresh food materials. Atmos. Environ. 89, 443e452. Klenbusch, M., 1986. Measurement of Gaseous Emission Rates from Land Surfaces Using an Emission-isolation Flux Chamber. User's guide. Radian Corp., Austin, TX (USA). Kuebler, L.S., Olsson, S.B., Weniger, R., Hansson, B.S., 2011. Neuronal processing of complex mixtures establishes a unique odor representation in the moth antennal lobe. Front. Neural Circ. 5, 7.
rrez-Ga
lvez, A., 2012. Signal and data processing for machine Marco, S., Gutie olfaction and chemical sensing: a review. Sens. J. IEEE 12 (11), 3189e3214. McGinley, M.A., McGinley, C.M., 2008. Odor threshold emission factors for common WWTP processes. Proc. Water Environ Feder 2008 (4), 530e546. ~ oz, R., Sivret, E.C., Parcsi, G., Lebrero, R., Wang, X., Suffet, I.H., Stuetz, R.M., 2010. Mun Monitoring techniques for odour abatement assessment. Water Res. 44 (18), 5129e5149. Nake, A., Dubreuil, B., Raynaud, C., Talou, T., 2005. Outdoor in situ monitoring of volatile emissions from wastewater treatment plants with two portable technologies of electronic noses. Sensor. Actuator. B Chem. 106 (1), 36e39.
, P., 2000. Using the clasNicolas, J., Romain, A.-C., Wiertz, V., Maternova, J., Andre sification model of an electronic nose to assign unknown malodours to environmental sources and to monitor them continuously. Sensor. Actuator. B Chem. 69 (3), 366e371. € € Gürek, A.G., Atilla, D., Ahsen, V., Ince, _ Oztürk, Z.Z., Tasaltin, C., Engin, G.O., M., 2009. Evaluation of a fast wastewater odour characterisation procedure using a chemical sensor array. Environ. Monit. Assess. 151 (1e4), 369. Pearce, T.C., Schiffman, S.S., Nagle, H.T., Gardner, J.W., 2003. Handbook of Machine Olfaction: Electronic Nose Technology. Wiley-VCH, Weinheim. Phaisangittisagul, E., Nagle, H.T., Areekul, V., 2010. Intelligent method for sensor subset selection for machine olfaction. Sensor. Actuator. B Chem. 145 (1), 507e515.
, T., Guy, C., 2007. Odor monitoring at the city of Montreal waste Purenne, P., Page water treatment plant. Water Practice 1 (2), 1e14. Qu, G., Omotoso, M.M., El-Din, M.G., Feddes, J.J., 2008. Development of an integrated sensor to measure odors. Environ. Monit. Assess. 144 (1e3), 277e283. Rajbansi, B., Sarkar, U., Hobbs, S.E., 2014. Hazardous odor markers from sewage wastewater: a step towards simultaneous assessment, dearomatization and removal. J Tai Inst Chem Eng 45 (4), 1549e1557. Romero-Flores, A., McConnell, L.L., Hapeman, C.J., Ramirez, M., Torrents, A., 2017. Evaluation of an electronic nose for odorant and process monitoring of alkalinestabilized biosolids production. Chemosphere 186, 151e159. Scott, S.M., James, D., Ali, Z., 2007. Data analysis for electronic nose systems. Microsc. Acta 156 (3e4), 183e207.
ntola, P., Del Rosso, R., 2007. Development of a system for the Sironi, S., Capelli, L., Ce continuous monitoring of odours from a composting plant: focus on training, data processing and results validation methods. Sensor. Actuator. B Chem. 124 (2), 336e346. Smith, S.W., 1999. The Scientist and Engineer's Guide to Digital Signal Processing. California Technical Publishing, San Diego, CA, pp. 1e32. Sohn, J.H., Hudson, N., Gallagher, E., Dunlop, M., Zeller, L., Atzeni, M., 2008. Implementation of an electronic nose for continuous odour monitoring in a poultry shed. Sensor. Actuator. B Chem. 133 (1), 60e69. Stuetz, R.M., Fenner, R.A., Engin, G., 1999a. Assessment of odours from sewage treatment works by an electronic nose, H 2 S analysis and olfactometry. Water Res. 33 (2), 453e461. Stuetz, R.M., Fenner, R.A., Engin, G., 1999b. Characterisation of wastewater using an electronic nose. Water Res. 33 (2), 442e452. Stuetz, R.M., Nicolas, J., 2001. Sensor arrays: an inspired idea or an objective measurement of environmental odours? Water Sci. Technol. 44 (9), 53e58. Sun, J., Hu, S., Sharma, K.R., Keller-Lehmann, B., Yuan, Z., 2014. An efficient method for measuring dissolved VOSCs in wastewater using GCeSCD with static headspace technique. Water Res. 52, 208e217. ~ o de un registrador de odorantes porta
til basado en Valiente Romero, R., 2014. Disen
n de concentracio
n. Ingeniero en Automa
tica. sensores MOS y modulacio
cnico Jose
Antonio Echeverría, La Habana, p. 129. Instituto Superior Polite Wilson, A.D., 2012. Review of electronic-nose technologies and algorithms to detect hazardous chemicals in the environment. Proc. Technol. 1, 453e463. €stro €m, M., Eriksson, L., 2001. PLS-regression: a basic tool of chemoWold, S., Sjo metrics. Chemometr. Intell. Lab. Syst. 58 (2), 109e130. Zarra, T., Naddeo, V., Belgiorno, V., Reiser, M., Kranert, M., 2009. Instrumental characterization of odour: a combination of olfactory and analytical methods. Water Sci. Technol. 59 (8), 1603e1609. Zarra, T., Reiser, M., Naddeo, V., Belgiorno, V., Kranert, M., 2014. Odour emissions characterization from wastewater treatment plants by different measurement methods. Chem. Eng. Trans. 40, 37e42.