On the Temporal Stability of Analyte Recognition with an E-Nose Based on a Metal Oxide Sensor Array in Practical Applications

Abstract

The paper deals with a functional instability of electronic nose (e-nose) units which significantly limits their real-life applications. Here we demonstrate how to approach this issue with example of an e-nose based on a metal oxide sensor array developed at the Karlsruhe Institute of Technology (Germany). We consider the instability of e-nose operation at different time scales ranging from minutes to many years. To test the e-nose we employ open-air and headspace sampling of analyte odors. The multivariate recognition algorithm to process the multisensor array signals is based on the linear discriminant analysis method. Accounting for the received results, we argue that the stability of device operation is mostly affected by accidental changes in the ambient air composition. To overcome instabilities, we introduce the add-training procedure which is found to successfully manage both the temporal changes of ambient and the drift of multisensor array properties, even long-term. The method can be easily implemented in practical applications of e-noses and improve prospects for device marketing.

Keywords: ambient air; electronic nose; honey recognition; instability; linear discriminant analysis; long-term stability; meat quality control.

Figure 1

The e-nose data tested at the short-term scale versus elementary LDA model. The ellipses indicate the positions of training class data obtained at elementary model mode. (a) The position of analyte-related ellipses in the LDA elementary model; (b) two rounds of analyte recognition by the E-nose. “a” and “b” designate the test e-nose data transitions at the exposures to apple of first sort and honey, respectively. The e-nose data measured at round 1 (1a and 1b) have been taken about 15 min after the collection of training sampling and within 15 min one after another. Then, after a pause of 40 min, the e-nose data have been collected at round 2 (2a and 2b).

Figure 2

(a) General view of the class ellipses in LDA space illustrating the e-nose model built to recognize apples of different kinds (“Borovinka”, “Golden Delicious” and “Kirgizskoe Zimnee”) and walnuts (“Ak-Terek”) of different freshness; (b) Enlarged area framed in the plot (a) to draw e-nose data trajectories at the LDA space during exposures to Kirgizskoe Zimnee apple at the first (1) and second (2) rounds following LDA model building. The e-nose data of 2nd round are more than two times farther away from the training class position than these of 1st round.

Figure 3

The results of LDA processing of the e-nose data to discriminate meat freshness (a) and honeys (b). The curves in the background show typical transitions of the resistance of exemplary sensor segment of the array during the exposures. The centers of LDA model classes are marked by asterisks symbols and the confidence intervals of 95% probability for each class are indicated with vertical bars. The meat freshness recognition models are numbered as: 1–model of the first series; 2–the second series one; 3–the combined model of both the series. The honey classes’ recognition models: 1–the samples of different harvesting times, 2–the samples of different harvesting places.

Figure 4

The time evolution of resistance of the exemplary sensor segment (#33) of the e-nose array measured in the mornings during the second measurement series (blue points) versus the day. Dashed yellow line indicates rather steep linear resistance trend. Solid green line discloses the step-like character of resistance curve under approximation by polynomial function of sixth degree. The dash-dotted red line marks the resistance level measured during the first series.

Figure 5

The S_TOM(N) and S_TRM(N) functions related to e-nose measurements concerning the recognition of meat freshness (a) and honey origin (b). N is number of days of e-nose measurements at the two series: (a) 1–6 are first series, 7–24 are the second series; (b) 1–7 are first series, 1–8 are the second series. At a point with the abscissa N, the training pool includes N days [34,36]. The S_TOM(N) gives the LOO scores for this training pool. The S_TRM(N) gives the recognition scores of all the involved E-nose measurements taken after the Nth day.

Figure 6

The photo of the e-nose installed in Karlsruhe Institute of Technology to demonstrate the performance of analyte recognition. Next to the e-Nose, one can see a mounted multisensor array chips and the sputter mask to fabricate it. The analytes to recognize are the odor emitted by felt pen and fuel gas emitted by lighter. The photo has been taken when the e-nose is exposed to the pen.

Figure 7

A typical R(t) behavior of the array segments in the e-nose during an event of its exposure to fuel gas emitted by a lighter. The resistance curves of different segments are marked with different colors. The arrow points the moment of gas appearance.

Figure 8

The PC screened image of the e-nose software under the long-term operation. Several last measurement points are displayed with the blue crosses while the last point is marked by the red color. Blue crosses and the classification table at the right side show a transition of e-nose state from the “ambient air” (“Luft”) to the “odor induced by felt-pen” (“Farbstift”). The violet crosses indicate e-nose signals which are classified as “others” (“unknown”). The red bar fraction shown in the right plot corresponds linearly to the position of red cross: it disappears when the cross is on the ellipse’s boundary and takes the complete bar, if the cross is in the center. The class “lighter” is marked as “Feuerzeug”.

Figure 9

LDA diagrams illustrating e-nose recognition of three deviating exposures to analytes. (a)-recognition with the LDA model built on the training pool which contains data of 35 previous exposures: the recognition score is 0.48. Inclusion of only one of the deviating e-nose data into the training pool of the LDA model radically improves the recognition of the two other deviating exposures: (b–d) show the three possible variants of the inclusions; the scores are 0.8 in (b) and 1.0 in (c,d).

Figure 10

Distributions of the segment resistance in the e-nose array (cf. [23]) recorded under ambient air during the period of four years, 2009–2013. Since the maximum measurable resistance of the sensors in the array of the e-nose is 55 MΩ, not all the sensors are plotted here. The positions of the median values taken over the array are encircled. The profiles are chosen to correspond to the similar season (spring).

Figure 11

The long-term evolution of median sensor resistance of the array in the e-nose measured from 2009–2017. The median evolution expresses the year-season and short-term variations. The averaged resistance represents the mean value of few typical sensor segments #22–25 from the array; points correspond to the distributions given in Figure 10 to be fit by curve built by their polynomial approximation. The averaged points illustrate the long-term drift and the eventual stabilization of sensor resistances. Oscillations of the curve to be seen since the fourth quarter of 2015 are reasoned by jumping of the position of median from one segment to another.

Figure 12

The Fourier spectrum of the median resistance value of the multisensor array in the e-nose measured in the years 2009–2017. The spectrum sections are given near the (a) annual peak; (b) weekly peak; and (c) daily peak. Yearly and daily peaks exceed essentially the local mean-spectrum level; the weekly peak does not.

Figure 13

Evolution of S_TRM(N) and the characteristic distance of the analyte-related clusters in the LDA space, D, with increasing the number N of e-nose exposures included into the training pool. The data of e-nose exposures to odors of pen and lighter have been recorded between April 2009 and September 2013. The distance D is normalized by the value observed at the initial e-nose training.