Understanding the role of sensor diversity and redundancy to encode for chemical information in gas sensor arrays

Abstract

[eng] Electronic noses (e-noses) have been utilized during the past three decades as general purpose instruments for chemical sensing. These instruments are inspired by natural olfactory systems, where fine odour discrimination is performed without the necessity for highly specialized receptors. Instead, odour information is extracted in these systems using arrays of broadly tuned receptors organized in a convergent pathway. Such a sensing architecture allows combining the responses of the array of receptors, giving rise to particular representations of the different odour stimuli. The key advantage provided by this approach is that odour representation is more efficient and robust when the encoding is performed by the population of receptors than by any of its individual elements (hyper-acuity). A population of receptors obtains its maximum performance in encoding odour stimulus features when it balances the benefits of sensory diversity and redundancy. By sensor diversity we understand the number of different receptor types responsible for enhancing the variability of the array response to a collection of odours. Likewise, by sensor redundancy we refer to the average number of receptor replicates on a population. The role of sensor redundancy accounts for the robustness to receptor damage and noise exhibited by the odour stimuli representation. This variety of odour receptor types along with its outstanding number of receptors is characteristic of natural olfactory systems. Though, traditional electronic noses tend to exhibit a limited number of sensor units with very much correlated responses to odour stimuli. Several strategies to enhance odour representation in gas sensor arrays are based on boosting sensor diversity and redundancy. However, it has not been until recently that large arrays of cross-selective have become technologically available. In this dissertation, we have developed one of these new generation arrays to investigate the advantages odour stimuli representation through population coding in artificial olfaction. In particular, we proposed to build a chemical sensing system based on an array of metal oxide (MOX) gas sensors, and endowed with a high a degree of sensor diversity and redundancy. We proposed the use this bio-inspired sensing architecture alongside statistical pattern recognition techniques to cope with some of the unsolved problems in machine olfaction (robustness to sensor damage, feature selection, and calibration transfer). The main contributions of this work were the following: We defined functionally sensor diversity and redundancy. These definitions were based on the clustering of the array features according to their similitude when responding to an odour dataset. We compared the different manner how natural and artificial olfactory systems encode for odour information using simple sensors models. We found that natural olfactory system principally encoded odour information in terms of odour quality, whereas that artificial ones in terms of odour quantity. Also, we studied the effect of sensor noise on odour concentration encoding. We proposed to decrease the contribution of the sensor noise by means of the redundant sensor feature averaging and sensor array optimization. These strategies were effective in case of independent sensor noise, but not for removing common sources of sensor noise. Similarly, we detected the importance of sensor failure dependency on the odour discrimination capabilities of a sensor. We found that this sensor fault distribution across had to be independent of the sensor type to prevent a dramatic worsening on the array’s predictive performance. In addition to this, we proposed an update of a feature selection method including a dimensionality reduction stage so as to take into account the redundant information provided by the sensor array. Finally, we performed instrument standardization between temperature modulated sensor arrays to correct global shifts of temperature. A method to categorize the quality of the calibration transfer based on the bias-variance trade-off was presented.

[spa] La nariz electrónica (e-nose) ha sido utilizada durante las últimas tres décadas como instrumento de propósito general para la detección química. Este instrumento está inspirado en los sistemas olfativos naturales, donde la discriminación de olores se realiza eficientemente sin la necesidad de receptores altamente especializados. La ventaja clave proporcionada por esta aproximación es que la representación de los olores es más eficiente y robusta cuando la codificación del olor es llevada a cabo por una población de receptores, pues esta supera la calidad de cualquiera realizada por sus elementos individuales. Una población de receptores obtiene su máximo rendimiento en la codificación de las características de un estímulo odorífero cuando se equilibran los beneficios de la diversidad y redundancia sensoriales. Lamentablemente, las narices electrónicas tradicionales tienden a exhibir un número limitado de sensores con respuestas muy correlacionas ante diferentes conjuntos de estímulos odoríferos. Sin embargo, no ha sido hasta hace relativamente poco que la creación grandes matrices de sensores con selectividades cruzadas ha sido tecnológicamente posibles. El objetivo de esta tesis es el desarrollo de una de estas matrices de nueva generación para investigar las ventajas de la representación de los estímulos odoríferos través de una codificación poblacional soportada por la diversidad y redundancia sensoriales. En particular, hemos construido un sistema de detección química basado en una matriz de sensores de gas de óxido metálico (MOX), y dotada de un alto grado de diversidad y redundancia sensoriales. Hemos utilizado esta arquitectura de detección química bioinspirada junto técnicas de reconocimiento de patrones estadísticas para hacer frente a algunos de los problemas sin resolver en olfacción artificial, a saber, la robustez al fallo de sensores, la selección de características, y la transferencia de calibración entre diferentes narices electrónicas.