2016年4月8日星期五

Metal Oxide Semiconductor Gas Sensors and Neural Networks

Primitive mobile life forms, and the mobile cells of higher animals, derive their motivation and direct their navigation by chemical senses. In the simplest cases these creatures are hard-wired to swim toward nutrient concentration gradients, and to swim against irritant concentration gradients. The vestiges in humans of the sometimes extremely sensitive, selective, and differential chemical senses of primitive forms are taste, our ability to detect and identify four classes of chemicals in water solution on our tongues, and smell or olfaction, our ability to detect and identify many gases, vapors, and complex mixtures in the air passing through our noses.

Primitive "autonomous" mobile robots derive their motivation from human directive, and direct their navigation largely via artificial visual and acoustic senses. Their autonomy is thus in the discretion they are allowed in planning and executing getting from here to there; the heres and the theres generally are dictated more-or-less explicitly by humans. An evolutionary path for sophistication of the navigational abilities of mobile robots looks self-evident, and in fact such evolution is well underway.

On the other hand an evolutionary path along which mobile robots might even begin to progress toward replacing motivation by human fiat with self directed motivation is difficult to conceive, at least if the senses that drive motivation are to remain as complex as vision and hearing.

But self directed motivation of mobile robots based on one or more chemical senses is easy to imagine, and indeed is probably within the capability of existing technology. For example, we could build right now a mobile robot that would meander around a chemical plant, sniffing as it goes for gas leaks (or the vapors of liquid leaks), navigating toward them while avoiding hazards visually, yet always motivated as to overall direction (an undetermined endpoint notwithstanding) by the chemical concentration gradient, with end-point navigation perhaps directed to the offending pipe fissure or open valve by acoustic homing toward the source of the hiss, and effecting simple repairs or summoning human assistance (with appropriate hazard warning) based on the fusion of chemical,visual, and acoustic sensory information in context. Similar scenarios can easily be imagined for firefighting robots, prospecting robots, rescue robots, contraband nterception robots, and others.

Toward this end, my colleagues and I have been studying chemical sensing, not with the traditional goal of inventing alternative or improved instruments for doing precise chemical analysis, but rather from the perspective of exploring alternative approaches to an artificial sense of smell to motivate autonomous activity.

First, I describe the characteristics of metal oxide semiconductor gas sensors used as gas sensors, particularly for combustibles. Response, as a change in resistance, depends on combustible gas identity and concentration, sensor temperature, and also on the concentrations of other gases and vapors present,i.e.,response to a mixture is not a linear superposition of responses to the individual components. These sensors inherently respond to broad classes of compounds, but differential responses conducive to signature identification can be induced by variation of parameters such as temperature and catalyst nature and concentration. I then survey the olfactory system, in contrast with other biological sensory and sensor interpretation systems, in analogy with MOS gas sensors, and in analogy between its neurological organization and the organization of simple artificial neural networks.

I then describe the class of large area thick film MOS resistors that we have constructed with a spatial gradient in relative sensitivity induced by differential heating. Multiple resistance measurements along the temperature gradient yield a multi-feature signature that we analyze and interpret using artificial neural networks. In particular, I contrast the approach that emphasizes binary or nearly binary unit activities, and thus forces every mixture into a class distinct from the classes of its single species components, with an approach that is less rigidly binary, encouraging creation of only enough classes to account for the individual species present, and putting mixtures into multiple classes in proportion to the concentration of the mixture components in each class.




没有评论:

发表评论