Sensor Variability Helps Build a Better E-Nose

Thirty-two-sensor array reliably sniffs out differences between apples and pears

2 min read

Sensor Variability Helps Build a Better E-Nose

Diversity is a good thing in the never-ending effort to produce electronic noses that can compete with human and canine noses (and honey bee antennae) to smell out all kinds of important stuff:  finding contraband like explosives and narcotics, watching over the maturation of a vintage wine, gauging blood glucose levels without a pin-prick…even detecting the telltale biochemical signature of a malignant tumor.

A team from the Polytechnic University of Valencia in Spain and the University of Gävle in Sweden has rigged an array of 32 commercially available sensors that can sniff a bit of crushed fruit and tell if the source was an apple or a pear.

Valencia’s  José Pelegrí Sebastiá and collaborators used seven different models of off-the-shelf  metal-oxide semiconductor gas sensors. The sensors are marketed to detect a wide variety of gases, including propane, natural gas, carbon monoxide, alcohol, toluene, and a variety of others, by measuring  conductance changes across a narrow gap between two electrodes as trace gases adsorb to the electrode’s surface. Different gasses produce distinct curves of voltage over time, and each type of sensor has different response characteristics. There is, moreover, some sensor-to-sensor variation in response.  And the particular sensors chosen (from Figaro Engineering) include internal heaters: changing the sensor’s operating temperature further modifies its response curves.

The research group exploited these characteristics to mix sensor models and operating temperatures to produce an eNose, an array of 32 sensors, each with its own distinct response curve. A bit of aromatic air wafted into the eNose will send 32 channels of dynamic voltage data into a data recorder and on into a variety of pattern-recognition algorithms.

As a test, the researchers put bits of apples or pears into the sample chamber, let the fruit’s volatile compounds diffuse into the air for a moment, and then injected the “sniff” into the eNose.

The team characterized each of the 32 resulting signals by three dynamic parameters: transient slope (the rate of initial rapid increase when the gas is first detected), saturation slope (the slower rate of voltage increase as the senor reaches its greatest response), and maximum slope  (measured when the sensor is closed off from the sample).

The investigators subjected each of these parameters to principal component analysis, so that each electronic sniff generated patterns of 96 points in three-dimensional space for each of 20 sample runs. They then fed the results into 10 different pattern recognition programs* and asked each one, in effect, “Is it an apple, or is it a pear?”  All but one of them could correctly identify apples or pears nine times out of 10, or better. And one pattern-reader—IB1, a nearest-neighbor algorithm— got it right on the nose 100% of the time. This, the researchers say, is better than most people could do. (Dogs’ noses, of course, are a couple of orders of magnitude more sensitive than ours, so the eNose has something to shoot for.)

*The eNose group used the University of Waikato (New Zealand) open-source Weka data-mining library. The analyses included a standard naïve Bayesian Network, radial basis functions, linear logistic regression models, sequential minimum optimization, a nearest-neighbor algorithm, an entropic distance algorithm, a voting algorithm, a tree model, a nearest-neighbor algorithm that learns rules, and a partial decision tree model.

Photos: Apples and Pears-Gabor Izso/iStockphoto; eNose University of Gävle and Polytechnic University of Valencia

The Conversation (0)