Gas Identification System using Graded Temperature Sensor and Neural Net Interpretation

An innovative approach is described for enhancing the selectivity of an integrated multi-element thick film gas sensor. A temperature gradient maintained along the sensor surface induces spatial sensitivity and selectivity gradients. The response map of the sensor is examined in the vapor of several organic compounds and their binary mixtures. Subtle but recognizable and separable signatures are observed. Emphasis is on identification, since quantitation of identified mixtures is straightforward. An effective and robust classification technique using a neural network trained via the back propagation method is described. Introduction The demand for selectively sensing vapors and gases is keen. For example, it would be valuable to distinguish toxic exhaust gases from fuel vapors and warn of either of these in the passenger compartment of automobiles, boats, and airplanes. Similarly, it is important for a domestic gas leakage detection system to avoid false alarms by differentiating between the odors of alcoholic beverages and the utility gas. Additional examples that require selectively sensing gases and vapors are found in military, industrial, and domestic applications. Most chemical analysis systems for gaseous samples, such as gas chromatography, are expensive, complicated to use, and at best marginally portable. In contrast, relatively inexpensive, simple, highly portable gas sensing devices based on resistivity changes of semiconducting materials are in common use. One of the most common materials is tin oxide, SnO2. It has high sensitivity to a large number of gases and is relatively easy to use. However its high sensitivity to many gases is a weak point when it comes to sample discrimination and mixture analysis. Substantial efforts have been directed at improving the selectivity of SnO2 based sensors by various approaches using filters, special absorbing layers, and various catalysts or promoters. The bibliography contains an extensive survey of the SnO2 gas sensor literature [1,2, 3, 4, 5, 6, 7, 8, 9,10] [11,12,13, 14, 15,16,17, 18,19, 20, 21, 22, 23]. Temperature programming is among the many techniques that have proven useful for enhancing selectivity. The sensing mechanism of SnO2 to reducing agents is believed to be due to the chemical reactions between their molecules and surface adsorbed and ionized oxygen as O~ and perhaps other species. There is an optimum temperature at which sensitivity maximizes. If the temperature is lower, the reaction is too slow to give high sensitivity; if the temperature is higher, the overall oxidation reaction proceeds too rapidly. In the latter case, diffusion of the reducing agent is confined to a thin layer near the surface, and the effective concentration of reducing agent seen by the sensor bulk is decreased. Thus if an SnO2 sensor is ramped through an appropriate temperature range, the sensitivity to a gaseous sample will show a peak at a particular temperature. Sensitivity peaks associated with the various gases present will generally appear at different temperatures in this temperature programming technique. However, hysteresis in the sensor resistance during temperature cycling generally broadens these peaks. Thus, temperature programming of the sensor is not a common technique used for selectivity enhancement; a notable exception is the Figaro Corporation approach to CO alarming, which employs a Taguchi sensor and a complex temperature programming method [19]. We are approaching discrimination by temperature from another direction. Instead of cycling the temperature of a single sensor, we apply a static temperature gradient along a linear sensor array. In fact, the "array" is a continuous film on a ceramic substrate, and the discrete array elements are defined only by the contact metalizatfon. Since the sensing elements on the substrates are fabricated under exactly the same conditions, their sensing properties should be extremely uniform at constant tenrperature. But when these sensors are at different temperatures, their relative responses to the components of gas mixtures differ, providing recognizable signatures. Significantly different response patterns of an SnO2 array to ethanol, methanol, heptane and their binary mixtures were observed using the temperature gradient technique. The reproducibility of these response patterns is acceptable for laboratory studies, but stability improvement is a subject of ongoing research toward field applications. The coarse spatial resolution of our sensor (seventeen elements in 23 mm), the difficulty in 1 An alternative approach, operating m conetanf temperature wfth a gmienf m catalyst concentration or composite, and a combination approach w ih peipendfeular temperature and catalyst gradients is also employed and wM be described in future repoiti. in fact, tome of our devices entpioy bot i approaches, m± pmpmdkwlm' graders in temperature and' catalyst measuring precisely the temperature distribution, and the small but non-negligible intrinsic structural differences among the sensing elements, makes the development of an analytical model to describe the sensing properties of the integrated thick film sensor very difficult. An alternative approach is to interpret pragmatically the experimental data via machine learning, pattern recognition, and perhaps knowledge based artificial intelligence methods. In this report we test this approach when implemented via the neural network [24] with training by back propagation method [25]. Sensors The sensor arrays we use in these experiments are fabricated on commercial alumina ceramic substrates intended for hybrid circuits. The area of the substrate is compatible with a 20-pin dual-inline integrated circuit. A gold electrode metallization pattern is printed, then fired at high temperature onto the substrate, as shown in Figure 1. Next a layer of a modified commercial SnO2 semiconductor ink is screen printed on top of the electrodes. The gold pattern provides stable electrical contact between the sensing layer and the measurement equipment. The modified SnO2 ink and the screen printing technique provide a cost-effective method to fabricate a high quality sensor array. Catalysts such as fine Pt metallic powder are blended into the commercial inks to enhance sensitivity. The organic binders in the ink are burned away by firing in an oven. A detailed description of the fabrication procedure, which is actually carried out by collaborators at Oak Ridge National Laboratories, is given in [26, 27,28].