Cheap, reliable analysis of a chemical environment would greatly benefit a wide range of application areas. For example, homeland security appeals for constant analysis of public spaces for evidence of explosives or toxic chemical agents. Other applications include industrial control, health diagnostics, and environmental monitoring. Chemical sensor arrays or ‘electronic noses’ are an up-and-coming technology that may provide for inexpensive, real-time chemical detection. These devices typically use optical- or electrical-sensing mechanisms that respond to a broad range of chemicals. Consequently, the devices must be trained before end-use by measuring the response on exposure to known substances.
A number of operational schemes and data-analysis techniques have been used to obtain as rich a data set as possible from sensor arrays and, then, transform the data into the identity of a chemical analyte (molecule of interest) using pattern-recognition schemes.1,2 These approaches are largely successful in ideal conditions but tend to fail when the sensors' responses change due to age or the presence of competing analytes. In addition, a similar chemical compound to one present in the training set is often not recognized as similar by the matrix-manipulation scheme. As an analogy, imagine learning the smell of wine by sniffing a glass of Syrah and then not recognizing that Pinot Noir is also wine.
The lack of ability to generalize knowledge from the set of training analytes arises because the very purpose of the traditional data-manipulation schemes is to find the features that most uniquely discriminate each analyte in the training set. Thus, there is a fundamental conflict between identifying an analyte and recognizing it as a member of a class: the latter requires finding data features common to a group of analytes. Recently,3 we developed a scheme using chemiresistive metal-oxide sensor arrays based on micro-hotplates to obtain a very large data set. We subsequently split the twin problems of class recognition and unique identification. The research was inspired by parallel work occurring at the National Institutes of Health examining animal olfaction. Animals seem to solve these problems by first recognizing the general class of an odor and then refining to a more specific identity over a period of time.
To start, we fabricated micro-hotplate chemical sensors as described in prior work.4 We used 16 total sensors, comprising two nominally identical copies of eight different sensing materials. Each sensor was, in turn, heated through a ramped temperature profile from 50 to 500°C using embedded micro-heaters. We measured the resistance of each sensor at every 1°C during the ramp. The use of different materials and temperatures created a very large analytical space. Each sensor-temperature combination was treated as an independent ‘neuron,’ though there was a large degree of correlation between many of them.
The sensors were then trained by measuring in nominally pure dry air with and without one of a range of simple organic analytes diluted to around 3 parts per million. Following accelerated aging, the sensors were again measured on exposure to some of the original analytes as well as a few new analytes not used in the training set. To address both the class recognition and unique identification problems, we built an analysis scheme that solved for analyte identity in a hierarchical fashion: see Figure 1. For each question in the hierarchy (e.g., Is this hydrocarbon an alkane or aromatic?), the region of data space that best answered that specific question was used.
Figure 1. The hierarchical classification scheme by which analyte identification was progressively resolved. ppm: Parts per million. ppb: Parts per billion.
During the validation portion of the experimentation, the hierarchical scheme was able to both correctly identify the vast majority of the analytes repeated from the training set and correctly categorize the new analytes. For example, after training with methanol and ethanol, 1-propanol and 2-propanol were correctly found to belong to the same class. As an added benefit, we found improved robustness to aging effects. As a further test, this sensing scheme was able to differentiate analyte concentration by adding concentration as a final step in the hierarchy.
The key advance in this work was the use of multiple recognition steps arranged to achieve progressively finer representations of the analyte, using optimal data for each question. Solving for the identity in one step requires that data beneficial for the identification of each and every analyte be used at the same time, even if some of that data is unproductive or highly error-prone for the specific analyte under test. While the technique was validated on a particular type of sensor, the concept is quite general and should be applicable to many electronic-nose technologies. Future work includes expanding this technique to identify and classify components from within more complex chemical mixtures and changing backgrounds.
Joshua L. Hertz
University of Delaware
1. P. C. Jurs, G. A. Bakken, H. E. McClelland, Computational methods for the analysis of chemical sensor array data from volatile analytes, Chem. Rev.
100, no. 7, pp. 2649-2678, 2000. doi:10.1021/cr9800964
2. R. Gutierrez-Osuna, Pattern analysis for machine olfaction: a review, IEEE Sensors J.
2, no. 3, pp. 189-202, 2002. doi:10.1109/JSEN.2002.800688
3. B. Raman, J. L. Hertz, K. D. Benkstein, S. Semancik, Bioinspired methodology for artificial olfaction, Anal. Chem
. 80, no. 22, pp. 8364-8371, 2008. doi:10.1021/ac8007048
4. S. Semancik, R. E. Cavicchi, M. C. Wheeler, J. E. Tiffany, G. E. Poirier, R. M. Walton, J. S. Suehle, B. Panchapakesan, D. L. DeVoe, Microhotplate platforms for chemical sensor research, Sensors Actuat. B
77, no. 1–2, pp. 579-591, 2001. doi:10.1016/S0925-4005(01)00695-5