sensors Article
Suppression of Strong Background Interference on E-Nose Sensors in an Open Country Environment Fengchun Tian 1, *, Jian Zhang 1 , Simon X. Yang 2 , Zhenzhen Zhao 1 , Zhifang Liang 1 , Yan Liu 1 and Di Wang 1 1
2
*
College of Communication Engineering, Chongqing University, 174 Sha Pingba, Chongqing 400044, China;
[email protected] (J.Z.);
[email protected] (Z.Z.);
[email protected] (Z.L.);
[email protected] (Y.L.);
[email protected] (D.W.) Advanced Robotics and Intelligent Systems (ARIS) Lab, School of Engineering, University of Guelph, Guelph, ON N1G 2W1, Canada;
[email protected] Correspondence:
[email protected]; Tel./Fax: +86-23-6511-1745
Academic Editor: W. Rudolf Seitz Received: 16 November 2015; Accepted: 5 February 2016; Published: 16 February 2016
Abstract: The feature extraction technique for an electronic nose (e-nose) applied in tobacco smell detection in an open country/outdoor environment with periodic background strong interference is studied in this paper. Principal component analysis (PCA), Independent component analysis (ICA), re-filtering and a priori knowledge are combined to separate and suppress background interference on the e-nose. By the coefficient of multiple correlation (CMC), it can be verified that a better separation of environmental temperature, humidity, and atmospheric pressure variation related background interference factors can be obtained with ICA. By re-filtering according to the on-site interference characteristics a composite smell curve was obtained which is more related to true smell information based on the tobacco curer’s experience. Keywords: electronic nose sensors; outdoor; background interference; suppression
1. Introduction Electronic nose (e-nose) techniques are being more and more widely used in areas such as food evaluation, medical diagnosis, environmental monitoring and industrial control [1], but the interference from environmental temperature, humidity, atmospheric pressure variation and other non-target smells is still a bottleneck which frustrates people and limits the development of e-nose technology. Two factors cause this situation. On the one hand, the key part of the e-nose, i.e., most metal oxide semiconductor (MOS) gas sensors, are not perfect and their responses are easily affected by temperature, humidity and atmospheric pressure variations. On the other hand, the e-nose sensors are inherently susceptible to interference since each gas sensor of the e-nose should have cross sensitivity, i.e., it should respond to more than one gas. This is either an advantage or a disadvantage. The benefit of cross sensitivity is that an e-nose may measure many kinds of smells with a limited number of sensors, while various interferences will result in responses in the e-nose sensors. The interferences resulting from other gases and environmental factors cannot be effectively separated either by traditional filtering or by wavelet transforms since these interferences are almost entirely mixed with the target gases, which are to be detected, both in the frequency and time domain, so solving the problem of environmental interference suppression in an e-nose is an important task. Currently the main methods of compensating for environmental temperature and humidity variations are methods based on temperature and humidity compensation models, artificial neural networks (ANNs), PCA combined with ANNs, etc. For example, a compensation method based on knowledge modification was proposed in [2], where the environmental temperature and humidity Sensors 2016, 16, 233; doi:10.3390/s16020233
www.mdpi.com/journal/sensors
Sensors 2016, 16, 233
2 of 17
were used as inputs of an ANN to build the correction model. The ANN and the automatic Bayesian regularization were used to form an approximate regression function retaining good generalization properties to reduce the influence of the environment on an e-nose in [3]. In [4] the dimensions of sample data consisting of temperature, humidity and gas sensor responses were reduced by PCA, then features were chosen according to Wilks’ rule, and finally the compensation to ambient temperature and humidity was realized. A thermistor was adopted to compensate for the influence of temperature-incurred changes of MOS gas sensor responses in [5], while the effect of humidity was ignored. Two equal sensors were used to minimize the interference of temperature with one sensor for signal measurement and another as reference, the difference of the two sensors was used as the output signal in [6]. The responses of temperature and humidity sensors together with that of gas sensors were used as inputs of an ANN to realize compensation of temperature and humidity influences in [7,8]. The information of temperature and humidity sensors was input to a data merging center consisting of wavelet ANNs to perform compensation to environmental temperature and humidity changes in [9]. A physical way was used to compensate the temperature and humidity influence in [10], where SiO2 was used as the sensing material with a titanium thermistor being used to maintain a constant temperature. Combined with data standardization, PCA and ANN, the gas concentration prediction was realized with QCM as gas sensor [11]. Information of gas sensors under various temperature and humidity conditions was collected to calculate the corresponding drift coefficients, and a compensation was realized in [12]. After the sensor response signals were analyzed by ICA, the correlation of each independent component with temperature and humidity was calculated to determine and remove the temperature and humidity factors [13]. Meanwhile, the background interference correction of e-noses can be carried out by methods based on correlation, ICA, ANN as well as support vector machine (SVM). Instead of modifying the system hardware, these are software compensation methods. For example, to perform wavelet transform (WT) on sensor signals followed by calculating of spatial correlation coefficients of WT between smells of infected and healthy mice under the same WT scale, the background interference was minimized according to correlation coefficients [14]. Zhang et al. divided the anti-interference process into two steps: (1) determination of interference, (2) correction of the interfered signal. The gases were partitioned into two categories: target and non-target (i.e., those with the exception of the target). The response of each sensor was categorized first by least squared SVM (LSSVM), back propagation (BP) network and was replaced by the most nearby signal if it was judged as an interference [15]. Al-Maskari et al. adopted methods based on kernel Fuzzy C-Mean clustering and Fuzzy SVM to increase the sorting accuracy of e-nose data with noise and drift [16]. The sensor array output was analyzed by ICA with the responses to background interference being used to construct a reference vector, and the correlation coefficients between the components of ICA and the reference vector were calculated to determine and delete background interferences [17]. Feng et al. suppressed interference by removing the components of response matrix orthogonal to those of the target matrix with RFB parameters optimized by particle swarm optimization (PSO) [18]. In summary, various methods of environmental interference have been proposed, however, they were either used in the recognition/categorization of smells, or only suitable to the application scenarios where background interferences as only their temperature and humidity effects were compensated. We propose a method for effectively separating environment temperature, humidity, atmospheric pressure variation and reducing non-target smell interference. The remainder of the paper is arranged as follows: a brief overview of the tobacco curing process is introduced in Section 2. The developed system for collecting the smell, the analysis of the interference existing in the system and the proposed method to restrain the interference are presented in Section 3. The result of our method, i.e., the tobacco smell curve, is presented in Section 4. Then the data certification using CMC to validate the effectiveness of the proposed method and some discussions are provided in Section 5. Finally, conclusions are given in Section 6.
Sensors 2016, 16, 233
3 of 17
Sensors 2016, 16, 233
3 of 17
2. Tobacco Curing Curing Process—A Process—A Brief Brief Overview Overview 2. Tobacco The The curer’s curer’s chart chart for for tobacco tobacco curing curing is is illustrated illustrated in in Figure Figure 1. 1. The The three-stage-curing three-stage-curing craft craft was was adopted in our experiments, which is common in tobacco curing factories. The curing adopted in our experiments, which is common in tobacco curing factories. The curing chart chart is is divided divided into into 19 19 stages stages in in detail detail or or three three coarse coarse stages stages (i.e., (i.e., yellowing yellowing stage, stage, color-fixing color-fixing stage stage and and stem-drying stem-drying stage). stage).
Figure 1. Three-stage-curing chart.
A flue-curing flue-curing process temperature and humidity in A process may may last last six sixto toseven sevendays days(144–168 (144–168h). h).The The temperature and humidity the curing barn are reflected by dry-bulb temperature and wet-bulb temperature (psychrometer), in the curing barn are reflected by dry-bulb temperature and wet-bulb temperature (psychrometer), respectively.The Thered redline lineininFigure Figure1 1represents represents dry-bulb temperature denotes respectively. thethe dry-bulb temperature andand thethe blueblue oneone denotes the the wet-bulb temperature. The temperature and humidity changed during the whole curing process. wet-bulb temperature. The temperature and humidity changed during the whole curing process. 3. Experiments 3. Experiments 3.1. Developed E-Nose System for Tobacco Tobacco Flue-Curing Flue-Curing With enzymes, thousands of chemical components are produced and they With the thecatalysis catalysisofofvarious various enzymes, thousands of chemical components are produced and release their corresponding smellssmells duringduring tobacco curingcuring [19]. Professionals may determine the curing they release their corresponding tobacco [19]. Professionals may determine the stage quality of tobacco by sniffing smells. e-nose simulate the olfaction function curingand stage and quality of tobacco by these sniffing theseAn smells. Anmay e-nose may simulate the olfaction of humanofbeings, and it is expected find out to thefind features andfeatures rules ofand tobacco variation by function human beings, and it istoexpected out the rulessmell of tobacco smell utilizing sothe as to provide a clue for automatic control of tobacco curing. variationthe by e-nose utilizing e-nose so as to provide a clue for automatic control of tobacco curing. Here low price and simplicity are the key points to be considered since this is a widely used application in tobacco factories, so a simple and specialized e-nose, rather than a commercialized common purpose e-nose for lab experiments, was adopted in our system. The specialized e-nose system designed for tobacco curing smell feature extraction is illustrated in Figure 2. The e-nose system comprises an air filter, sensor array, rotameter, vacuum micro-pump, data acquisition card and IEEE 485 bus, etc. The chemical components of tobacco which have important contributions to their smells can be mainly categorized as phenols, organic acids, lipid substances, aldehyde substances, alcohol substances, etc. [19,20], so a set of sensors including nine MOS gas sensors (cf., Table 1) and temperature, humidity, atmospheric pressure sensors were selected to detect smells during the tobacco flue-curing process.
Figure 2. E-nose system designed for smell features extracting in tobacco curing.
3
With the catalysis of various enzymes, thousands of chemical components are produced and they release their corresponding smells during tobacco curing [19]. Professionals may determine the curing stage and quality of tobacco by sniffing these smells. An e-nose may simulate the olfaction function beings, and it is expected to find out the features and rules of tobacco 4smell Sensors 2016,of16,human 233 of 17 variation by utilizing the e-nose so as to provide a clue for automatic control of tobacco curing.
Figure 2. E-nose system designed for smell features extracting in tobacco curing. Table 1. Sensors used in the E-nose. Many of the below sensors have responses to alcohol, but their responses to these key chemicals are different among suppliers, providing an increased amount of 3 chemical information. Sensor Type
No.
Related Sensitivity
Manufacturers
TGS826 TGS813 TGS822 TGS 2600
1 2 3 4
FIGARO, Osaka, Japan FIGARO, Osaka, Japan FIGARO, Osaka, Japan FIGARO, Osaka, Japan
TGS 2602
5
MQ135
6
MQ138
7
WSP2111 SP3S-AQ2 MPX4100AP DS600 HIH4000
8 9 10 11 12
Isobutane, ethanol, ammonia, hydrogen Methane, propane, isobutane Ethanol, organic solvents Cigarette smoke Volatile Organic Compounds (VOCs), ammonia, hydrogen sulfide Ammonia, sulfide, BTEX, acetone, toluene, ethanol, carbon monoxide Alcohols, ketones, aldehydes, aromatics, organic solvents Toluene, benzene, ethanol, acetone VOC, hydrogen, ethanol, methane, ammonia Atmospheric pressure Temperature Humidity
FIGARO, Osaka, Japan Winsen, Zhengzhou, China Winsen, Zhengzhou, China Winsen, Zhengzhou, China FIS, Hyogo, Japan Freescale, Austen, TX, USA MAXIM, Sunnyvale, CA, USA Honeywell, Morristown, NJ, USA
Since ambient temperature, humidity and atmospheric pressure (T/H/P) have strong influences on the response of gas sensors, corresponding T/H/P sensors were used in the sensor array of the e-nose so as a compensation can be made in the subsequent algorithm. The outputs of the sensor array were amplified by the signal regulation board and converted to data by an A/D converter card with a 12 bits capacity. Then, these data were transmitted to the PC of the monitoring center via an IEEE 488 bus. To avoid the problem of re-pollution, the vacuum pump was connected to the outlet of the chamber housing the sensor array. The rotameter was used to maintain a constant flow of gas. Photos of the designed e-nose system and experimental site (a tobacco curing barn in the countryside) are shown in Figures 3 and 4 respectively.
Sensors 2016, 16, 233
5 of 17
Sensors 2016, 16, 233 Sensors 2016, 16, 233
5 of 17 5 of 17
Figure 3. Photo of the e-nose system.
The whole e-nose was put beside the tobacco barn, and a micro pump was used to pump gas Photo of the the e-nose system. Figure 3. Photo of into the chamber from the barn, as shown in Figure 2.e-nose The whole e-nose was put beside the tobacco barn, and a micro pump was used to pump gas into the chamber from the barn, as shown in Figure 2.
Figure 4. Photo of the tobacco curing barn. Figure 4. Photo of the tobacco curing barn.
Since the distance between the bulk curing-barn and the monitoring center PC ranges was Thetens whole e-nose the wasIEEE put Figure beside tobacco barn, andcuring adata micro pump was used to pump gas into 4.the Photo of the tobacco barn. several of meters, 485 bus was used for remote transmission. The operating process the chamber from the barn, as shown in Figure 2. of the e-nose is as follows. The three-way valve was first switched to the outlet of the air filter so Since the distance distancebetween betweenthethe bulk curing-barn and monitoring the monitoring center PC ranges was bulk curing-barn and centerfor PC15ranges was several purified airthe was pumped to the sensor array chamber bythe the vacuum pump min. During this several tens of meters, the485 IEEE bus wasfor used for remote data transmission. The operating process tens of meters, IEEE bus485 used remote data Thewas operating process the period, the gas the sensors worked inwas baseline status. Then, thetransmission. three-way valve switched to theofvent of the e-nose is as follows. The three-way valve was first switched to theofoutlet offilter the air filter so e-nose is as follows. The three-way valve was first switched to the outlet the air so purified of the curing-barn and the gas, which contained the smells to be measured, was pumped to the sensor purified air was pumped to thearray sensor array chamber by the vacuum pump for 15 min. During this air was pumped sensor by the vacuum pump for 15 sensor min. During this period, array chamber forto10the min. After that,chamber the purified air was pumped into the array chamber for period, gas sensors worked in baseline status. Then, the three-way valve was switched to theofvent the gas the sensors in baseline status. Then, the three-way valvepump was switched to the the 15 min to purgeworked the sensors. Finally, it took 20 min for the vacuum to rest and onevent cycle was of the curing-barn and thewhich gas, which contained the smells to be measured, was pumped to the sensor curing-barn and the contained the smells was pumped finished as shown ingas, Figure 5 where the response oftoa be gasmeasured, sensor during the cycletoisthe alsosensor given.array The array chamber for 10 min. After that, the purified air was pumped into the sensor array chamber for chamber for 10was min.repeated After that, the purified air was pumped intoprocess. the sensor array chamber for 15 min same process during the whole tobacco curing The repeated cycles lasted 15 min tothe purge the sensors. tookfor 20 min for thepump vacuum pump toone rest andwas one cycle was to purge Finally, itFinally, took 20itmin vacuum to rest cycle as 7 days since sensors. the whole curing process lasts 7 the days. Each hour, there is and 1 sampling periodfinished (10 min), finished asFigure shown in Figure 5response where the response of aduring gas sensor during thegiven. cycle is also given. The shown in 5 where the of a gas sensor the cycle is also The same process and 20 samples are collected during this 10-minute period. Since it needs 7 days to finish a whole same process during was repeated during the whole tobacco process. The lasted repeateddays cycles lasted was repeated whole process.=curing The the curing process, almostthe 7 (days) ×tobacco 24 (h) ×curing 20 (samples) 3360repeated samples cycles are collected.7All thesince samples 7whole days curing since the wholelasts curing process lasts 7 days. Each hour, there is 1 sampling period (10 min), process 7 days. Each hour, there is 1 sampling period (10 min), and 20 samples are concatenated together to make a data sequence (curve). and 20 samples are collected duringperiod. this 10-minute to curing finish aprocess, whole are collected during this 10-minute Since it period. needs 7Since days ittoneeds finish7adays whole curing process, almost 7 (days) × 24 (h) × 20 (samples) = 3360 samples are collected. All the samples 5 are concatenated together to make a data sequence (curve). 5
Sensors 2016, 16, 233
6 of 17
almost 7 (days) ˆ 24 (h) ˆ 20 (samples) = 3360 samples are collected. All the samples are concatenated Sensors 2016, 16, 233 6 of 17 Sensors 2016, 233 a data sequence (curve). 6 of 17 together to16, make
Figure 5. Smell data acquisition cycle and corresponding sensor response response cycle. Figure 5. Smell data acquisition cycle and corresponding sensor response cycle. cycle.
3.2. Data Data Analysis—Strong Analysis—Strong Background Background Interference Interference Existing Existing at at Outdoor Outdoor Environment Environment 3.2. Interference Existing at Outdoor Environment The interference interference to to an an outdoor outdoor e-nose e-nose lies lies mainly mainly in in variations variations of of temperature, temperature, humidity, The temperature, humidity, humidity, atmospheric pressure and background smell. atmospheric pressure pressure and and background background smell. smell. Interference from Temperature, Humidity Humidity and andAtmospheric AtmosphericPressure Pressure 3.2.1. 3.2.1. Interference from Temperature, Temperature, Humidity and Atmospheric Pressure The gas gas sensors used ininour our e-nose areare very susceptive to temperature temperature and humidity humidity (T/H).(T/H). Their gas sensors sensorsused usedin our e-nose very susceptive to temperature and humidity The e-nose are very susceptive to and (T/H). Their baselines vary vary greatly withwith T/H,T/H, eveneven if only only carrier gas/purified airairappears appears [21]. The Their baselines greatly if only carrier gas/purified appears[21]. [21]. The The actual baselines vary greatly with T/H, even if carrier gas/purified air actual temperature, humidity and atmospheric pressure of the environment collected by corresponding temperature, humidity and atmospheric pressure of the environment environment collected collected by corresponding corresponding sensors in inthe thesensor sensorchamber chamberare are shown in Figures Figures 6–8, respectively. The sensor sensor array chamber is sensors shown in Figures 6–86–8, respectively. The sensor array array chamber is made sensors in the sensor chamber are shown in respectively. The chamber is made of stainless steel with a layer of Teflon coating and it is not heated, so the temperature in the of stainless steel with a layer of Teflon coating and it is not heated, so the temperature in the chamber made of stainless steel with a layer of Teflon coating and it is not heated, so the temperature in the chamber shown in Figure wasboth affected both by the the temperature temperature of the the gasbarn fromand the the barn and the air air shown in shown Figure in 6 was affected by the temperature of the gas of from airand outdoors. chamber Figure 66 was affected both by the gas from the barn the ˝ outdoors. Since the thetemperature outdoor temperature temperature difference betweenand daytime and was nighttime was more more than Since the outdoor difference difference between daytime nighttime more than 10 Cthan (the outdoors. Since outdoor between daytime and nighttime was 10 °C (the seven peaks and seven valleys correspond to 7 days’ daytime and nighttime, respectively), seven peaks and seven valleys correspond to 7 days’ daytime and nighttime, respectively), obviously, 10 °C (the seven peaks and seven valleys correspond to 7 days’ daytime and nighttime, respectively), obviously, theofinfluence influence of environmental environmental T/H/P may not be be ignored ignored in aatobacco whole tobacco tobacco curing process. process. the influence environmental T/H/P may not may be ignored in a whole curing process. obviously, the of T/H/P not in whole curing
Figure 6. Variation of environment temperature. temperature. Figure 6. 6. Variation Variation of Figure of environment environment temperature.
The humidity humidity level level during during the the whole whole curing curing process process is is shown shown in in Figure Figure 7. 7. To To get get the the purified purified air, air, The a set of filters such as activated carbon, molecular sieves and some canisters were used to filter out a set of filters such as activated carbon, molecular sieves and some canisters were used to filter out CO, SO SO22 and and water, water, do do the the level level of of humidity humidity was was low low at at the the purging purging stage. stage. However, However, the the humidity humidity CO, level in in the the chamber chamber dynamically dynamically changed changed during during the the whole whole curing curing process process due due to to the the water water released released level from tobacco. from tobacco. 6
Sensors 2016, 2016, 16, 233 233 Sensors Sensors 2016, 16, 16, 233
7 of 17 7 of 17
Figure 7. 7. Variation Variation of of humidity humidity level. level. Figure Figure 7. Variation of humidity level.
Figure 8. Variation of atmospheric pressure. Figure8. 8.Variation Variation of of atmospheric atmospheric pressure. pressure. Figure
3.2.2. Interference Interference of Environmental Environmental Smells Smells on on the the Sensor Sensor Array. Array. 3.2.2. The humidityoflevel during the whole curing process is shown in Figure 7. To get the purified air, Due to the cross sensitivity of e-nose gas sensors, they are susceptible to many smells. In the the a setDue of filters such as sensitivity activated carbon, molecular sieves andare some canistersto were used to filter out to the cross of e-nose gas sensors, they susceptible many smells. In scenario of current application, the interference comes mainly from pollutant pollutant gases, such such ashumidity SO22,, CO CO CO, SO2of and water,application, do the levelthe of interference humidity was low mainly at the purging stage. However, theas scenario current comes from gases, SO and nitrogen oxides, etc. which were produced by the burning coal. The TGS813 ( FIGARO, Osaka, level in the chamber dynamically changed during the whole curing process due to the water released and nitrogen oxides, etc. which were produced by the burning coal. The TGS813 (FIGARO, Osaka, Japan is one one sensor sensor of of the the e-nose e-nose sensor sensor array. array. The The on-site on-site collected collected response response of of the the TGS813 TGS813 and and from))tobacco. Japan is corresponding amplitude of its Fourier transform are shown in Figure 9 where those points of sensor corresponding amplitude of its Fourier transform are shown in Figure 9 where those points of sensor 3.2.2. Interference of Environmental Smells on the Sensor Array purging, bump resting resting and baseline baseline are are all omitted. omitted. To reduce reduce the influence influence of of environmental environmental factors, factors, purging, bump and all To the each Due pointtoin inthe thecross curve was obtained obtained by subtracting subtracting the baseline and divided divided by the thesmells. difference of sensitivity of e-nose gas sensors, they are susceptible to many In the each point the curve was by the baseline and by difference of maximum and minimum of sensor response. The sampling frequency of the data acquisition is set to scenario ofand current application, the response. interference mainly from pollutant gases, such as SO maximum minimum of sensor Thecomes sampling frequency of the data acquisition is set to 2 , CO s = 1/30 Hz = 0.033 Hz. Though the response of only one sensor is shown here, the responses of the nitrogen oxides, etc. which the were produced by the burning The here, TGS813 ffand s = 1/30 Hz = 0.033 Hz. Though response of only one sensor coal. is shown the(FIGARO, responses Osaka, of the other eight sensors are similar. Japan)eight is one sensor ofsimilar. the e-nose sensor array. The on-site collected response of the TGS813 and other sensors are There are 20 intervals (peaks) in Figure Figure 9b and andare it is isshown easy to toincalculate calculate the corresponding frequency, corresponding amplitude its Fourier transform Figure 9the where those pointsfrequency, of sensor There are 20 intervals of (peaks) in 9b it easy corresponding i.e., f s/20 = (1/30)/20 = 1/600 (Hz), so the corresponding time is 600 s (10 min). It corresponds to the purging, resting and baseline omitted. To reduce of environmental i.e., fs/20 =bump (1/30)/20 = 1/600 (Hz), so are the all corresponding time isthe 600influence s (10 min). It correspondsfactors, to the sampling period of e-nose in Figure 5, and can be seen obviously in Figure 7. each pointperiod in theofcurve was obtained by subtracting baseline divided by the difference of sampling e-nose in Figure 5, and can be seen the obviously inand Figure 7. maximum and minimum of sensor response. The sampling frequency of the data acquisition is set to fs = 1/30 Hz = 0.033 Hz. Though the response of only one sensor is shown here, the responses of the other eight sensors are similar.
77
Sensors 2016, 16, 233 Sensors 2016, 16, 233
8 of 17 8 of 17
Figure Figure 9. 9. Response Response (a) (a) and and its its spectrum spectrum (b) (b) of of e-nose e-nose sensor sensor TGS813. TGS813.
3.3. Background Interference Suppression Procedure in an E-Nose There are 20 intervals (peaks) in Figure 9b and it is easy to calculate the corresponding frequency, background an e-nose, a method of separating i.e., fsTo /20control = (1/30)/20 = 1/600interference (Hz), so theincorresponding time is 600 s (10 min). environment-related It corresponds to the interference factors under strong background interference was proposed in this paper. The steps of sampling period of e-nose in Figure 5, and can be seen obviously in Figure 7. the algorithm are as follows: 3.3. Background Interference Suppression Procedure in an E-Nose (1) Pre-processing (including baseline removal, low-pass filtering and standardization); To Principal control background in an e-nose, a method of separating environment-related (2) componentinterference analysis (PCA); interference factors under strong background interference in this paper. Theas steps of the (3) Independent component analysis (ICA) with the was firstproposed several PCA components inputs of algorithm are as follows: ICA, removing those ICA output components which are strongly related to environmental interferences, then the ICA component which is more low-pass related tofiltering the true and smell signal, may be obtained; (1) Pre-processing (including baseline removal, standardization); (4) By re-filtering the above obtained ICA component, a useful signal with environmental and (2) Principal component analysis (PCA); background interferences suppressed in some is first obtained. (3) Independent component analysis (ICA) extent with the several PCA components as inputs of ICA, removing ICA output components are strongly environmental interferences, In thethose above analysis, some a prioriwhich information, suchrelated as thetotime of coal feeding (which then the ICA component which is more related to the true smell signal, may be obtained; corresponds to the time the pseudo-periodical interference gas is produced), ambient temperature, (4) Byand re-filtering the above obtained ICA component, a useful signal with environmental and humidity atmospheric pressure information (an approximately periodical change of temperature background interferences suppressed in some extent is obtained. and humidity is induced by the difference between daytime and nighttime) as well as the experts’
knowledge tobacco smell, some may bea used. In the on above analysis, priori information, such as the time of coal feeding (which corresponds to the time the pseudo-periodical interference gas is produced), ambient temperature, 3.3.1. Pre-Processing humidity and atmospheric pressure information (an approximately periodical change of temperature and humidity is induced by the the sensor difference between daytime and nighttime) as well as the experts’ This includes removing baseline, low-pass filtering and data standardization: knowledge on tobacco smell, may be used. (a) Baseline Removal and Normalization 3.3.1.APre-Processing typical gas sensor response curve is shown in Figure 5. It comprises the stages of baseline recovery, sampling, purgingthe and pump rest. To lessenfiltering the influence environment changes This includes removing sensor baseline, low-pass and dataofstandardization: (such as temperature, humidity and atmospheric pressure) on gas sensors, Equation (1) is used in (a) Baseline Removal and Normalization the standardization: A typical gas sensor response curve is shown si iniFigure 5. It comprises the stages of baseline xi To recovery, sampling, purging and pump rest. the influence of environment changes (1) lessen i (such as temperature, humidity and atmospheric pressure) on gas sensors, Equation (1) is used in the standardization: where si is the response (the voltage of the sensor, which is from a bleeder circuit [21]) of the ith s ´ µi xi “ i (1) sensor, μi is the baseline, σi is the standard deviation, xi is the output after standardization. In this σ way, the influence of ambient factors is expected to bei reduced by some degree. where si is the response (the voltage of the sensor, which is from a bleeder circuit [21]) of the ith sensor, µi is (b) the Low-Pass baseline, σFiltering i is the standard deviation, xi is the output after standardization. In this way, the influence of ambient factors is expected to be by some degree. The low-pass filter is used to filter out thereduced interference of white noise. The main parameters of a
low-pass filter include 3 dB bandwidth, cut-off frequency, type of filter, etc. By Fourier transform, the spectrum features and inherent interference frequencies may be found. For example, the frequency component corresponding to the 10 min sampling period may be found from Figure 9b, and the frequency bandwidth of the low-pass filter was selected to be able to filter out these periodical 8
Sensors 2016, 16, 233
9 of 17
(b) Low-Pass Filtering The low-pass filter is used to filter out the interference of white noise. The main parameters of a low-pass filter include 3 dB bandwidth, cut-off frequency, type of filter, etc. By Fourier transform, the spectrum features and inherent interference frequencies may be found. For example, the frequency Sensors 2016, 16,corresponding 233 9 of 17 component to the 10 min sampling period may be found from Figure 9b, and the frequency bandwidth of the low-pass filter was selected to be able to filter out these periodical interferences. The Thelow-pass low-passfilter filter a three low-pass filter. The frequency cut-off frequency of this interferences. is aisthree orderorder low-pass filter. The cut-off of this low-pass low-pass filter was set to 0.02 f s . The cut-off frequency of this low-pass filter should not be too low, filter was set to 0.02 fs . The cut-off frequency of this low-pass filter should not be too low, otherwise, otherwise, the independency of signal will be destroyed a bad ICA result mayThe occur. The the independency of signal source willsource be destroyed and a badand ICA result may occur. filtered filtered responses of the array of nine sensors are shown in Figure 10. responses of the array of nine sensors are shown in Figure 10.
Figure 10. Low-pass filtered filtered responses Figure 10. Low-pass responses of of the the nine nine gas gas sensors. sensors.
It is shown in Figure 10 that the peaks in these curves are very consistent though the amplitudes of the the gas gas sensors sensorsare aredifferent. different.Furthermore, Furthermore,it itis is found that appearance time of these peaks is found that thethe appearance time of these peaks is the the very the furnace, hintsthese that peaks these peaks were produced by the coal very timetime coal coal was was fed tofed thetofurnace, whichwhich hints that were produced by the coal smoke smoke entering the curing barn.each Sincetime eachwhen time coal when coal was fed, aamount large amount of pollution gases entering the curing barn. Since was fed, a large of pollution gases (such (such SOetc.), 2, CO, etc.), could whichbecould easily by smelled human was due produced due to as SO2as , CO, which easilybe smelled humanby noses, wasnoses, produced to incomplete incomplete burning. Thesewere pollutants intobarn the curing barnexhaust via its wet exhaust burning. These pollutants pouredwere intopoured the curing via its wet because it is because close to it is close to smokestacks. each of the e-nose sensors was sensitive to kinds some smokestacks. Though eachThough of the e-nose sensors was designed to bedesigned sensitivetotobe some specific specific of gases, they had extremely strong responses these smokes due to intensity their immense of gases,kinds they had extremely strong responses to these smokestodue to their immense while intensity while truewas tobacco smell wasbyoverwhelmed by these filters with true tobacco smell overwhelmed these interfering gases.interfering Low-passgases. filters Low-pass with different cut-off different cut-off frequencies were used, but the true signal of tobacco smell still could not be extracted frequencies were used, but the true signal of tobacco smell still could not be extracted since the true since the true in the same as that of smoke, environmental temperature, signal is in thesignal same is frequency bandfrequency as that ofband smoke, environmental temperature, humidity and humidity and atmospheric pressureNext, interferences. Next, ICAwith are the used to deal with atmospheric pressure interferences. the PCA and ICAthe arePCA usedand to deal data. the data. 3.3.2. PCA Model and ICA Model 3.3.2.APCA Model andresponse ICA Model 12-dimension was obtained from the array of 12 sensors. While the PCA was used to reduce dimension andresponse denoise, was the environment related factors were foundWhile out bythe checking the used PCA A 12-dimension obtained from the array of 12 sensors. PCA was load coefficients. to reduce dimension and denoise, the environment related factors were found out by checking the a useful method for blind source separation. It can separate statistically independent PCAICA loadiscoefficients. sources so method long as atfor most onesource sourceseparation. is of Gaussian distribution, as shown inindependent Figure 11. ICAeffectively is a useful blind It can separate statistically sources effectively so long as at most one source is of Gaussian distribution, as shown in Figure 11.
Figure 11. ICA model.
Assume there are n independent source signals s1(t), s2(t),…, sn(t) with their discrete forms s1,
A 12-dimension response was obtained from the array of 12 sensors. While the PCA was used to reduce dimension and denoise, the environment related factors were found out by checking the PCA load coefficients. ICA is a useful method for blind source separation. It can separate statistically independent Sensors 2016, 16, 233 10 of 17 sources effectively so long as at most one source is of Gaussian distribution, as shown in Figure 11.
Figure Figure 11. 11. ICA model.
Assume there there are arennindependent independentsource sourcesignals signalss1s(t), 1(t),s s(t), 2(t),…, forms ss11,, Assume . . . , ssnn(t) (t)with with their their discrete discrete forms 2 x1,xx1 ,2,…,x arguments observed, the purpose of ICA is toisestimate the ss22,…, , . . . s,n,snrespectively; , respectively; x2 , . m. .are ,xmthe aremthe m arguments observed, the purpose of ICA to estimate n independent source through the m measured arguments. The ICA model is given by: the n independent source through the m measured arguments. The ICA model is given by: (2)
X “9AS
where X = (x1 , x2 , . . . ,xm )T , S = (s1 , s2 , . . . , sn )T , A = (aij ), 1 ď i ď m; 1 ď j ď n. Both A and S are unknown. The only a priori information is that each component of S is statistically independent and at most one si (t) is of Gaussian distribution. The target of ICA is to estimate the separation matrix W, ˆ the optimal estimation of S, with the components of Sˆ which is the inverse of A, and further to get S, are independent as much as possible: Sˆ “ WX
(3)
The Fast ICA algorithm is a widely used algorithm and was adopted here [22]; it is based on the principle of maximized non-Gaussian feature. It searches out the non-Gaussian feature maximum of WX by iteration with the negative entropy approximation as target function. Herein the actual application scenario is as follows: m = 2, n = 2 and x1 , x2 , . . . ,xm are the output components of PCA (PCA1, PCA2), respectively; s1 (with ICA1 expressing its estimation) represents the source which is introduced by various gas smells; s2 denotes the other source (with ICA2 expressing its estimation) incurred by those environmental factors other than gas smells, such as temperature, humidity and atmospheric pressure. After pre-processing in the way given in Section 3.1, the outputs of the 12 sensors x1 , x2 , . . . , x12 were sent to PCA. The eigenvalues, accumulated contribution rate and coefficients of the first two PCA components y1 , y2 (denoted by PCA1 and PCA, respectively) are given in Tables 2 and 3. From Table 2, it can be found that the accumulated contribution rate of the first two components of PCA (i.e., PCA1, PCA2) reached 89.3%, so it is reasonable to use these two components as features of the sensor array. Table 2. PCA results. No.
Eigenvalue
Contribution Rate
Accumulated Contribution Rate
1 2 3 4 5 6 7 8 9 10 11 12
8.4115 2.3058 0.5775 0.3452 0.2226 0.0600 0.0399 0.0177 0.0092 0.0048 0.0031 0.0026
70.0955 19.2146 4.8127 2.8769 1.8553 0.5001 0.3324 0.1478 0.0768 0.0403 0.0258 0.0219
70.0955 89.3101 94.1227 96.9996 98.8549 99.3550 99.6874 99.8352 99.9119 99.9523 99.9781 100.0000
Sensors 2016, 16, 233
10 of 17
X AS
(2)
where X = (x1, x2,…,xm)T, S = (s1, s2,…, sn)T, A = (aij), 1 ≤ i ≤ m; 1 ≤ j ≤ n. Both A and S are unknown. The Sensors 2016, 16, 233 11 of 17 only a priori information is that each component of S is statistically independent and at most one si(t) is of Gaussian distribution. The target of ICA is to estimate the separation matrix W, which is the inverse of A, and furtherTable to get estimation S, with the components of Sˆ are Sˆ , the optimal 3. Coefficients of the first two PCAof components. independent as much as possible: No.
Inputs of PCA
PCA1
Sˆ = W 0.2990 X
PCA2
(3)
1 x1 ´0.1481 2 is a widely x2used algorithm 0.3211 0.0339here [22]; it is based on the The Fast ICA algorithm and was adopted 3 x3 0.3314 ´0.0858 principle of maximized non-Gaussian feature. It searches out the non-Gaussian feature maximum of 4 x4 0.3399 ´0.0350 WX by iteration with the 5negative entropy x5 approximation 0.3076as target function. ´0.0114 Herein the actual application scenario is as follows: m = 2, n = ´0.0212 2 and x1, x2,…,xm are the output 6 x6 0.3399 7 PCA2), respectively; x7 0.3378 ´0.0418its estimation) represents components of PCA (PCA1, s1 (with ICA1 expressing 8 x8 0.3302s2 denotes ´0.1376 the source which is introduced by various gas smells; the other source (with ICA2 9 x9 0.3304 ´0.1080 expressing its estimation) other than gas smells, such as 10 incurred byx10those environmental 0.1305 factors 0.5962 Sensors 2016, 16, humidity 233 11 of 17 temperature, and atmospheric pressure. 11 x11 0.1430 0.5390 x12 0.5381 After pre-processing12in the way given in Section0.0461 3.1, the outputs of the 12 sensors x1, x2, …, x12
Table it can be seen that the ty valuescontribution of PCA2 forrate atmospheric pressure, temperature were From sent to PCA.3, The eigenvalues, accumulated and coefficients of the first two and From humidity are 0.5962, 0.5390, 0.5381, respectively, and they are far bigger than those for other PCA components y 1 , y 2 (denoted by PCA1 and PCA, respectively) are given in Tables 2 and 3. From Table 3, it can be seen that the ty values of PCA2 for atmospheric pressure, temperature and factors. This means that PCA2 mainly reflects the influence of environmental factors while PCA1 Table 2, it are can0.5962, be found that 0.5381, the accumulated contribution ratefarofbigger the first two components PCA humidity 0.5390, respectively, and they are than those for otherof factors. reflects the PCA2) smell function of smell. two components of PCA are shown in as Figure 12. of the (i.e., PCA1, reached 89.3%, soThe itthe isfirst reasonable use these two factors components features This means that PCA2 mainly reflects influence oftoenvironmental while PCA1 reflects the sensor array. of smell. The first two components of PCA are shown in Figure 12. smell function Table 3. Coefficients of the first two PCA components.
No. Inputs of PCA PCA1 PCA2 1 x1 0.2990 −0.1481 2 x2 0.3211 0.0339 3 x3 0.3314 −0.0858 4 x4 0.3399 −0.0350 5 x5 0.3076 −0.0114 6 x6 0.3399 −0.0212 7 x7 0.3378 −0.0418 8 x8 0.3302 −0.1376 9 x9 0.3304 −0.1080 10 x10 0.1305 0.5962 11 x11 0.1430 0.5390 12 x 12 0.0461 0.5381 Figure of Figure 12. 12. Curves Curves of of the the first first and and second second components components of PCA. PCA. To further compensate the influenceTable of environmental 2. PCA results.factors, ICA was used after PCA. The Fast To further compensate the influence of environmental factors, ICA was used after PCA. The Fast ˆ in Equation (3)) are ICA algorithm for ICA was adopted. The outputs of ICA, i.e., ICA1 and ICA2 ( S ICA algorithm forEigenvalue ICA was adopted. The outputs of ICA, i.e., ICA1 andContribution ICA2 (Sˆ in Equation (3)) are No. Contribution Rate Accumulated Rate shown in 1Figure 13.8.4115 70.0955 70.0955 shown in Figure 13. 2 3 4 5 6 7 8 9 10 11 12
2.3058 0.5775 0.3452 0.2226 0.0600 0.0399 0.0177 0.0092 0.0048 0.0031 0.0026
19.2146 4.8127 2.8769 1.8553 0.5001 0.3324 0.1478 0.0768 0.0403 0.0258 0.0219
89.3101 94.1227 96.9996 98.8549 99.3550 99.6874 99.8352 99.9119 99.9523 99.9781 100.0000
10 Figure Figure 13. 13. Two Two outputs outputs (ICA1, (ICA1, ICA2) ICA2) of of ICA. ICA.
3.3.3. Second Low-Filtering Although ICA1 may be approximately regarded as being removed of the influence of environmental atmospheric pressure, temperature and humidity, there still exists (pseudo-periodical) interference from burning coal pollution. This has been confirmed by analyzing the response curves
Sensors 2016, 16, 233
12 of 17
3.3.3. Second Low-Filtering Although ICA1 may be approximately regarded as being removed of the influence of environmental atmospheric pressure, temperature and humidity, there still exists (pseudo-periodical) interference from burning coal pollution. This has been confirmed by analyzing the response curves of sensors in Figure 10, the consistency between the peaks of ICA1 curve and the time of coal feeding. Since the independency of various gas smell sources cannot be guaranteed, the true tobacco smell cannot be obtained by ICA only. To restrain the interference from burning coal (particularly from the early burning stage) and other noises, the second low-pass filtering (re-filtering) was used to filter the output of ICA. The cut-off frequency of this low-pass filter was set to be lower than the corresponding frequency of the pseudo periods and the interference from burning coal smoke was suppressed to a certain extent. It is worth noting that two low-pass filtering passes were adopted. The first low-pass filtering Sensors 2016, 16, 233 12 of 17 is before PCA, which is for white noise suppression; whereas the second low-pass filtering is after ICA, for removing burning interference. The cut-offfrequency frequencyofofthese thesetwo two filters filters is whichwhich is forisremoving the the burning coalcoal interference. The cut-off is different. of the the first first filtering filtering, different. The The cut-off cut-off frequency frequency of filtering is is set set far far bigger bigger than than that that of of the the second second filtering, otherwise, could notnot be separated correctly by ICA small otherwise, environmental environmentalinterference interferencesources sources could be separated correctly by since ICA too since too asmall cut-off frequency destroys independencies among sources. a cut-off frequency destroys independencies among sources. The (i.e., ICA1 afterafter the second low-pass filtering)filtering) together together with the curer’s evaluation The smell smellcurve curve (i.e., ICA1 the second low-pass with the curer’s on smells are shown in Figure 14. As there were strong interferences produced by the burning evaluation on smells are shown in Figure 14. As there were strong interferences produced bycoal, the some interference removal measures were taken. The interferences were not totally eliminated, but the burning coal, some interference removal measures were taken. The interferences were not totally curve shownbut in Figure 14 can match experiences the curer and some research it can be eliminated, the curve shown inthe Figure 14 can of match the experiences of the results, curer and some considered as a reference and approximation of the true smell variations. research results, it can be considered as a reference and approximation of the true smell variations.
Figure 14. 14. ICA1 low-pass re-filtering re-filtering (more (more related related to to tobacco tobacco smell). smell). Figure ICA1 after after low-pass
4. Result-Tobacco 4. Result-Tobacco Smell Smell Curve Curve The whole whole flue-curing flue-curing process process can can be be divided divided into into three three stages, stages, i.e., i.e., yellowing, yellowing, color-fixing color-fixing and and The stem-drying shown in Figure 14 based on the status of the tobacco. Meanwhile, a professional curer stem-drying shown in Figure 14 based on the status of the tobacco. Meanwhile, a professional curer in the the tobacco tobacco factory factory was was assigned assigned to to sniff sniff the the smell smell every every 33 h h and and his his experience experience was was used used as as an an in intelligent reference toto the experiences of the curer andand [19],[19], at the intelligent reference in in our oursmell smellmodelling. modelling.According According the experiences of the curer at yellowing stage, the smell is mixed grass flavor and it reduces gradually with the tobacco color the yellowing stage, the smell is mixed grass flavor and it reduces gradually with the tobacco color becoming more moreand andmore moreyellow. yellow.Meanwhile, Meanwhile, gradually sends a mellow flavor (aroma) becoming it it gradually sends outout a mellow flavor (aroma) andand the the smell curve begins to go up. During the color-fixing stage, after the leaves become totally yellow, the aroma reduces. Then at the early stem-drying stage, the dehydration of tobacco is stronger, the smell becomes a little bit pungent and it goes away along with the decreasing amount of moisture. When the free moisture of tobacco is exhausted, the smell is gradually replaced by the inherent smell of tobacco. The pungent smell is stronger at the end of flue-curing process. This curve is mainly
Sensors 2016, 16, 233
13 of 17
smell curve begins to go up. During the color-fixing stage, after the leaves become totally yellow, the aroma reduces. Then at the early stem-drying stage, the dehydration of tobacco is stronger, the smell becomes a little bit pungent and it goes away along with the decreasing amount of moisture. When the free moisture of tobacco is exhausted, the smell is gradually replaced by the inherent smell of tobacco. The pungent smell is stronger at the end of flue-curing process. This curve is mainly consistent with the curers’ olfactory perception, while it is only a qualitative reflection of the smell variation during the whole curing stage. From Figure 14, it can be found that the e-nose cannot differentiate between aromatic smell and pungent smell. Also there are some fluctuations of smell curve during the whole curing process and the exact mechanism has not been found yet. 5. Discussion We used coefficient of multiple correlation to measure the effectiveness of interference suppression in our method. 5.1. Coefficient of Multiple Correlation (CMC) CMC and its improved counterpart have been used in frequency or wavelet domain for pattern recognition of complex data set [23], elbow arthrosis and gait dynamic data analysis [24,25] etc. CMC is used to measure the correlation between variable y and other multiple variables x1 , x2 , . . . , xk . Bigger CMC means stronger correlation [26]. To quantify the validity of separating interferences by either PCA or ICA, we used CMC to measure the linear correlation between PCA/ICA component and the environmental factors. Bigger CMC means that the corresponding PCA/ICA component has stronger correlation with environmental factors. The steps of CMC computation are as follows: ˆ the regression of y to x1 , x2 , ..., xk : (1) Calculate y, yˆ “ β o ` β 1 x1 ` ... ` β k xk
(4)
(2) Calculate the CMC between y and x1 , x2 , ..., xk : ř R “ bř
py ´ yqpyˆ ´ yq ř py ´ yq2 pyˆ ´ yq2
(5)
5.2. Data Experiment Using CMC When computing the CMC of PCA, y should be the first component (PCA1) and second component (PCA2), respectively. Since only the correlation between those environmental factors (responses of atmospheric pressure, temperature and humidity sensors, i.e., x10 , x11 , x12 ) and PCA1/PCA2 is considered, only β0 and x10 , x11 , x12 appear in the right of Equation (4). Similarly, when the CMC of ICA is calculated, y should be the first component (ICA1) and second component (ICA2), respectively, while only β0 and x10 , x11 , x12 appear in the right of Equation (4). The regression (yˆ in Equation (4)) of the first and second components of PCA, i.e., PCA1 and PCA2 (y in Equation (5)) to environmental variables are given in Figures 15 and 16 respectively. It is obvious that the regression of PCA2 to yˆ is much better than that of PCA1. The regression coefficients calculated according to Equation (4) and the coefficient of multiple correlation R in Equation (5) are shown in Table 4. It is found that the CMC of PCA2 to environmental variables R is 0.9430, which is much bigger than that of PCA1, 0.4250. That means PCA2 is more related to environmental factors (atmospheric pressure, temperature, humidity).
obvious that the regression of PCA2 to yˆ is much better than that of PCA1. The regression coefficients calculated according to Equation (4) and the coefficient of multiple correlation R in Equation (5) are shown in Table 4. It is found that the CMC of PCA2 to environmental variables R is 0.9430, which is much bigger than that of PCA1, 0.4250. That means PCA2 is more related to Sensors 2016, 16, 233 14 of 17 environmental factors (atmospheric pressure, temperature, humidity).
Sensors 2016, 16, 233
14 of 17
Figure Regression of Figure 15. 15. Regression of PCA1 PCA1 to to environmental environmental factors. factors.
13
Figure Figure 16. 16. Regression Regression of of PCA2 PCA2 to to environmental environmental factors. factors. Table and regression coefficients of theoffirst of PCA to variables Table4.4.CMC CMC and regression coefficients thetwo firstcomponents two components of environmental PCA to environmental xvariables . 10 , x11 , x12 x10, x11, x12.
Component Component of PCA of PCA β0 PCA1 PCA1 PCA2 PCA2
β0 β10 β10 β11β11 −4.2102 −0.0541 0.1647 ´4.2102 ´0.0541 0.1647 0.0264 0.0274 0.0274 ´5.1636−5.1636 0.0264
ββ1212 R R 0.0021 0.4250 0.0021 0.4250 0.1653 0.1653 0.9430 0.9430
The CMC and regression coefficients of ICA components to environmental variables computed The CMC and regression coefficients of ICA components to environmental variables computed according to Equations (4) and (5) are shown in Table 5. Figures 17 and 18 are the regression curve of according to Equations (4) and (5) are shown in Table 5. Figures 17 and 18 are the regression curve of ICA1 and ICA2 to environmental variables, respectively. It can be found from Table 5 that since the ICA1 and ICA2 to environmental variables, respectively. It can be found from Table 5 that since the CMC of ICA2 to environmental variables (R = 0.9967) is far bigger than that of ICA1 (R = 0.2763), ICA2 CMC of ICA2 to environmental variables (R = 0.9967) is far bigger than that of ICA1 (R = 0.2763), ICA2 should be more related to environmental factors, and this can also be found from Figures 17 and 18 should be more related to environmental factors, and this can also be found from Figures 17 and 18 where the regression curve of the former is worse than the latter. Comparing Table 4 with Table 5, it where the regression curve of the former is worse than the latter. Comparing Table 4 with Table 5, it can be found that ICA2 is more related to environmental factors than PCA2 since the CMC of ICA2 can be found that ICA2 is more related to environmental factors than PCA2 since the CMC of ICA2 (0.9967) is bigger than that of PCA2 (0.9430). This is also confirmed by the fact that the curve fitting (0.9967) is bigger than that of PCA2 (0.9430). This is also confirmed by the fact that the curve fitting in in Figure 18 is better than that in Figure 16. Figure 18 is better than that in Figure 16. Table 5. CMC and regression coefficients of ICA components to environmental variables.
Component of ICA ICA1 ICA2
β0 −0.0437 3.6978
β10 0.0241 −0.0084
β11 −0.0447 −0.0394
β12 0.0433 −0.0999
R 0.2763 0.9967
CMC of ICA2 to environmental variables (R = 0.9967) is far bigger than that of ICA1 (R = 0.2763), ICA2 should be more related to environmental factors, and this can also be found from Figures 17 and 18 where the regression curve of the former is worse than the latter. Comparing Table 4 with Table 5, it can be found that ICA2 is more related to environmental factors than PCA2 since the CMC of ICA2 (0.9967) is bigger than that of PCA2 (0.9430). This is also confirmed by the fact that the curve fitting Sensors 2016, 16, 233 15 of 17 in Figure 18 is better than that in Figure 16. Table 5. CMC regression coefficients coefficients of of ICA ICA components components to to environmental environmental variables. variables. Table 5. CMC and and regression
Component of ICA β0 β10 β11 β0 β10 β11 ICA1 −0.0437 0.0241 −0.0447 ICA1 ICA2 ´0.0437 3.6978 0.0241 −0.0084 ´0.0447 −0.0394
Component of ICA ICA2
Sensors 2016, 16, 233
3.6978
´0.0084
´0.0394
β12 β12 0.0433 0.0433 −0.0999
´0.0999
Figure 17. Regression of ICA1 to environmental variables. Figure 17. Regression of ICA1 to environmental variables.
R R 0.2763 0.2763 0.9967 0.9967
15 of 17
14
Figure Figure 18. Regression Regression of of ICA2 ICA2 to environmental variables.
In contrast, the theCMC CMCofofICA1 ICA1 environmental factors is 0.2763 which is smaller thanofthat of In contrast, to to environmental factors is 0.2763 which is smaller than that PCA1 PCA1 (0.4250). That environmental means environmental influences much compensated for ICA1, which by is (0.4250). That means influences are muchare compensated for ICA1, which is obtained obtained by ICA after PCA, than for PCA1. i.e., ICA1 is more eligible to represent true tobacco smell. ICA after PCA, than for PCA1. i.e., ICA1 is more eligible to represent true tobacco smell. Comparing Comparing ICA1 Figure 17 10, with Figure 10, that one the canpeaks find that the peaks appearance times of ICA1 of Figure 17 of with Figure one can find appearance times of both figures areboth the figures are the same and they are consistent with the coal feeding time. same and they are consistent with the coal feeding time. 6. Conclusions Conclusions In the outdoor outdoorenvironment, environment,variations variations atmospheric pressure, temperature humidity as ofof atmospheric pressure, temperature andand humidity as well well as interferences arelarge, veryand large, and ansensor e-nosearray sensor array isaffected. stronglyBased affected. Based on the as interferences are very an e-nose is strongly on the application application scenario of the outdoor tobacco curing environment, method of restraining/compensating scenario of the outdoor tobacco curing environment, a methodaof restraining/compensating strong strong interference for e-noses was proposed. The outputs of PCA were used as inputs of ICA. By interference for e-noses was proposed. The outputs of PCA were used as inputs of ICA. By comparing comparing CMC of PCAand outputs and ICAwith outputs with environmental variables, the effectiveness the CMC ofthe PCA outputs ICA outputs environmental variables, the effectiveness of ICA of ICA after PCA in separating environmental influence is confirmed. a priori after PCA in separating environmental influence is confirmed. Combined Combined with a prioriwith knowledge, knowledge, low-pass filtering werenoise usedand to strong restrain noise andinterference strong background two low-passtwo filtering steps were usedsteps to restrain background (the smell interference smellcoal). produced by burningof coal). advantage of this method lies in effectively produced by(the burning The advantage this The method lies in effectively separating the sensor separating sensor of smell frominterference those of environmental interference interferences factors. The responses ofthe smell fromresponses those of environmental factors. The environmental environmental interferences were suppressed to amethod great degree. hope, this method will not only were suppressed to a great degree. We hope, this will notWe only be effective for e-nose sensors be effective for e-nose sensors used in tobacco curing, but also will it be of reference value for e-noses used outdoors for other purposes. Since only the knowledge of coal feeding time (which was the time coal smog appeared) and human olfactory perception to both smog and tobacco smell was used as a priori information besides low-pass re-filtering, the interference from the background smell cannot be removed entirely. Here we only studied the off-line case. Next, we will study the on-line (real-time)
Sensors 2016, 16, 233
16 of 17
used in tobacco curing, but also will it be of reference value for e-noses used outdoors for other purposes. Since only the knowledge of coal feeding time (which was the time coal smog appeared) and human olfactory perception to both smog and tobacco smell was used as a priori information besides low-pass re-filtering, the interference from the background smell cannot be removed entirely. Here we only studied the off-line case. Next, we will study the on-line (real-time) application, the combination of deep learning, big data technique and non-linear method [27] to restrain background interference as much as possible. The data and code used in this paper may also be downloaded [28]. Acknowledgments: This work was supported by the Colleges and Universities’ Research Foundation for Ph.D. Program of China (20120191110023), Project of Chongqing Science & Technology talent cultivating (cstc2013kjrc-tdjs40008), 2013 Innovative Team Construction Project of Chongqing Universities (KJTD201331). Author Contributions: Fengchun Tian designed the algorithms, analyzed the experiment results and wrote the initial manuscript. Jian Zhang and Yan Liu did the experiments and collected the data. Simon X. Yang made instruction on the signal processing. Zhenzhen Zhao and Di Wang worked on experimental analysis and manuscript modification. Zhifang Liang worked on data transformation and part of the algorithm programming. Conflicts of Interest: The authors declare no conflict of interest.
References 1.
2. 3.
4. 5. 6.
7.
8.
9.
10.
11.
12.
Bhattacharyya, N.; Seth, S.; Tudu, B.; Tamuly, P.; Jana, A.; Ghosh, D.; Bandyopadhyay, R.; Bhuyan, M. Monitoring of black tea fermentation process using electronic nose. J. Food Eng. 2007, 80, 1146–1156. [CrossRef] Nenova, Z.; Dimchev, G. Compensation of the Impact of Disturbing Factors on Gas Sensor Characteristics. Acta Polytech. Hung. 2013, 10, 97–111. Vito, S.D.; Piga, M.; Martinotto, L.; Francia, G.D. CO, NO2 and NOx urban pollution monitoring with on-field calibrated electronic nose by automatic bayesian regularization. Sens. Actuators B Chem. 2009, 143, 182–191. [CrossRef] Tian, X.L.; Yin, Y.; Liu, H.J. Research on Artificial Olfactory Sensor Technology for Liquor Identification. Food Sci. 2004, 2, 29–32. General Information for TGS Sensors, Figaro Engineering. Available online: http://120.52.72.42/www. figarosensor.com/c3pr90ntcsf0/products/general.pdf (accessed on 14 February 2016). Ferrari, V.; Marioli, D.; Taroni, A.; Ranucci, E.; Ferruti, P. Development and application of mass sensors based on flexural resonances in alumina beams. IEEE Trans. Ultrason. Ferroelectr. Freq. Control Soc. 1996, 43, 601–608. [CrossRef] Gardner, J.W.; Hines, E.L.; Molinier, F.; Bartlett, P.N.; Mottram, T.T. Prediction of health of dairy cattle from breath samples using neural network with parametric model of dynamic response of array of semiconducting gas sensors. Sci. Meas. Technol. 1999, 146, 102–106. [CrossRef] Xu, X.L.; Qiu, J.N.; Chen, C. A study on local sensor fusion of wireless sensor networks based on the neural network. In Proceedings of the International Conference on Machine Learning and Cybernetics, Kunming, China, 12–15 July 2008. Shi, J.F.; Tang, H.B.; Gong, H.Y. Application of wavelet neural network and multi-sensor data fusion technique in intelligent sensor. In Proceedings of the 7th World Congress on Intelligent Control and Automation, Chongqing, China, 25–27 June 2008; pp. 1114–1117. Emadi, T.A.; Shafai, C.; Freund, M.S.; Thomson, D.J.; Jayas, D.S.; White, N.D.G. Development of a polymer-based gas sensor—Humidity and CO2 sensitivity. In Proceedings of the Microsystems and Nanoelectronics Research Conference, Ottawa, ON, Canada, 13–14 October 2009; pp. 112–115. ˙ A study on the development of a Mumyakmaz, B.; Özmen, A.; Ebeoglu, ˘ M.A.; Ta¸saltın, C.; Gürol, I. compensation method for humidity effect in QCM sensor responses. Sens. Actuators B Chem. 2010, 147, 277–282. [CrossRef] Kashwan, K.R.; Bhuyan, M. Robust electronic-nose system with temperature and humidity drift compensation for tea and spice flavour discrimination. In Proceedings of the Sensors and the International Conference on new Techniques in Pharmaceutical and Biomedical Research, Kuala Lumpur, Malaysia, 5–7 September 2005; pp. 154–158.
Sensors 2016, 16, 233
13. 14.
15. 16. 17.
18.
19. 20. 21. 22. 23. 24.
25. 26. 27.
28.
17 of 17
Natale, C.D.; Martinelli, E.; Amico, A.D. Counteraction of environmental disturbances of electronic nose data by independent component analysis. Sens. Actuators B Chem. 2002, 82, 158–165. [CrossRef] Feng, J.; Tian, F.; Yan, J.; He, Q.; Shen, Y.; Pan, L. A background elimination method based on wavelet transform in wound infection detection by electronic nose. Sens. Actuators B Chem. 2011, 157, 395–400. [CrossRef] Zhang, L.; Tian, F.; Dang, L.; Li, G.; Peng, X.; Yin, X.; Liu, S. A novel background interferences elimination method in electronic nose using pattern recognition. Sens. Actuators A Phys. 2013, 201, 254–263. [CrossRef] Al-Maskari, S.; Li, X.; Liu, Q. An Effective Approach to Handling Noise and Drift in Electronic Noses. Lect. Notes Comput. Sci. 2014, 8506, 223–230. Tian, F.; Yan, J.; Xu, S.; Feng, J.; He, Q.; Shen, Y.; Jia, P. Background Interference Elimination in Wound Infection Detection by Electronic Nose Based on Reference Vector-based Independent Component Analysis. Inf. Technol. J. 2012, 7, 850–858. [CrossRef] Feng, J.; Tian, F.; Jia, P.; He, Q.; Shen, Y.; Fan, S. Improving the performance of electronic nose for wound infection detection using orthogonal signal correction and particle swarm optimization. Sens. Rev. 2012, 34, 389–395. [CrossRef] Gong, C.; Zhou, Y.; Yang, H. Introduction for Three Stage Curing of Flue-Cured Tobacco; Science Press: Beijing, China, 2006. Mao, Y.A.; Liu, W.; Huang, J.G.; Lu, H.B.; Zhong, K.J. Comparison on complete character of volatile components in cigarette cut tobacco by electronic nose detection. Chem. Sens. 2007, 27, 36–42. General information for TGS Sensors, Figaro Group. Available online: http://www.datasheet-archive.com/ TGS826-datasheet.html (accessed on 14 February 2016). Download Fast ICA for Matlab/Octave. Available online: http://research.ics.aalto.fi/ica/fastica/code/ dlcode.shtml (accessed on 14 February 2016). Chau, T. A review of analytical techniques for gait data. Part 2: Neural network and wavelet methods. Gait Posture 2001, 2, 102–120. [CrossRef] Ferrari, A.; Cutti, A.G.; Cappello, A. A new formulation of the coefficient of multiple correlation to assess the similarity of waveforms measured synchronously by different motion analysis protocols. Gait Posture 2010, 31, 540–542. [CrossRef] [PubMed] Farana, R.; Irwin, G.; Jandacka, D.; Uchytil, J.; Mullineaux, D.R. Elbow joint variability for different hand positions of the round off in gymnastics. Hum. Mov. Sci. 2015, 39, 88–100. [CrossRef] [PubMed] Kutner, M.H.; Nachtsheim, C.; Neter, J. Applied Linear Statistical Models, 5th ed.; McGraw-Hill/Irwin: Chicago, IL, USA, 2004. Lentka, L.; Smulko, J.M.; Ionescu, R.; Granqvist, C.G.; Kish, L.B. Determination of Gas Mixture Components Using Fluctuation Enhanced Sensing and The LS-SVM Regression Algorithm. Metrol. Meas. Syst. 2015, 22, 341–350. [CrossRef] Data and code for “Restrain of Strong Background Interference on E-nose Sensors”. Available online: ftp://202.202.66.156/ (accessed on 14 February 2016). © 2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons by Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).
