Improving the Robustness of Odor Sensing Systems by Multivariate Signal Processing

Improving the Robustness of Odor Sensing Systems by Multivariate Signal Processing

Marta Padilla (University of Barcelona, Spain), Jordi Fonollosa (University of Barcelona, Spain) and Santiago Marco (University of Barcelona, Spain)
DOI: 10.4018/978-1-4666-2521-1.ch014
OnDemand PDF Download:
No Current Special Offers


Electronic noses or artificial olfaction systems based on chemical gas sensors present lack of robustness, a problem that is mainly technological and requires more research to improve fabrication processes and develop new technologies. However, statistical signal processing can help to mathematically reduce those unwanted effects on the sensors responses before the prediction step. In this chapter, the authors explore the concept of robustness in electronic nose instruments and the use of several multivariate signal processing techniques to deal with two specific problems related to such lack of robustness: time instability (drift) and the detection of a possible faulty sensor in the array. In particular, three different techniques that deal with drift problems are reviewed. These techniques address drift by correction of unwanted variance, by taking advantage of the characteristics of a three-way data arrangement, or by using a blind strategy to extract information with chemical meaning. Finally, a method based on principal component analysis is presented for fault detection, faulty sensor identification, and correction of a fault in a sensor array.
Chapter Preview


Persaud and Dodd proposed the first electronic nose (e-nose) in 1982. Electronic noses (e-noses) or Artificial Olfaction (AO) systems were very promising for many qualitative and quantitative applications in the nineties. Researchers expected to develop a device capable of recognizing odors and characterize them with attribute descriptors such fruity, grassy, earthy, malty, etc., just like human sense of smell does. Electronic noses were also designed to be an alternative to the use of other instruments for gas analysis or detection. At that time, e-noses were expected to present a set of very interesting characteristics such as being of small size, low cost, low power consumption, fast, easy to use and to provide a prediction of the human odor impression. Consequently, especially interesting applications for these instruments were on-field fast detection and/or quantification of target gases or volatiles and global evaluations of aroma, even replacing human sensory panels, which would considerably reduce the related costs. However, despite these many potential advantages, nowadays, almost 30 years after the first device, e-noses have not reached the initially expected success. The main reason lies in the sensing area of the instrument, which exhibits poor selectivity and bad stability. Moreover, these instruments cannot mimic olfaction, and thus, give a human odor impression with the actual technology.

Nowadays, several sensing technologies are available for AO systems, but typically, these instruments are based on chemical gas sensors. In particular, these types of sensors suffer from cross sensitivities, time instability, dependence on previous gas exposures, different responses among identical sensors, etc. Such shortcomings affect negatively the reproducibility and reliability of the final instrument, and thus, its robustness. Therefore, instruments based on these sensors suffer from lack of robustness and do not give enough reproducible results. Consequently, e-noses are not (yet) suitable for industry. The nature of the problems and the lack of robustness that chemical gas sensors show are mainly technological and affect sensors of all state of the art technologies, though to different degrees. In the future, those deficiencies will be mostly overcome by improving the fabrication process and by developing new technologies. However, while more research is made on these areas, statistical signal processing can help to mathematically compensate or reduce the effect of the mentioned issues before pattern recognition.

The aim of this chapter is to explore the concept of robustness in AO systems and the use of several statistical signal-processing techniques, along with sensor operation methods, that may improve the robustness of such AO systems. In particular, these techniques aim to correct or compensate the responses of the sensors affected by two specific problems related to the lack of robustness: sensor drift and failure of one or more sensors of an array. In order to improve time stability, three different signal-processing approaches recently proposed are reviewed. Such strategies make use of sensors time information either due to sampling transient or temperature modulation from different points of view. The first one uses the method Orthogonal Signal Correction to correct unwanted variance related to drift. The second approach proposes the use of an unsupervised Multiway technique (PARAFAC) on sensors transient signals to take advantage of the characteristics of a three way data arrangement. The last approach consists of the use of Multivariate Curve Resolution Alternating Least Squares (MCR-ALS) on temperature modulated metal oxide (MOX) sensors. MCR-ALS is a blind deconvolution strategy that extracts information with special meaning without any calibration step, by introducing prior knowledge into the data matrix decomposition through constraints. Finally, a method based on Principal Component Analysis (PCA) is presented for fault detection, faulty sensor identification, and correction in an array of chemical gas sensors. For this, a subtle fault based on the change on sensitivity produced by sensor poisoning is considered. Indeed, the detection of change in sensor sensitivity is a relevant problem that has received poor attention by the AO community.

Complete Chapter List

Search this Book: