Abstract
In recent years, odor recognition systems have been studied very actively. However, this field is difficult to research since there are a lot of kinds of odors, and they have not been characterized well. Because of these difficulties, odor property can be revealed from inspection such as sensory test performed by human. As the odor inspection by human fundamentally has fluctuation based on physical or climatic condition, an artificial odor recognition system is required in many industries.
The authors proposed a neural network called “Fuzzy Learning Vector Quantization (FLVQ)”, and applied it in the odor recognition system to identifying odor kinds and expressing sensory quantity obtained from the human sensory test.
FLVQ network receives the output pattern from the sensors, and is trained by the result of the human sensory test, such as a triangle test for whiskeys and a semantic differential (SD) method for flavors. The network can output odor quantity after learning. It has been confirmed from the outputs of the estimated sensory quantity that FLVQ has satisfactory ability to express human odor quantity.
Moreover, adsorption membranes for odor sensors were selected to improve the odor recognition system. The authors proposed three statistical methods for membrane selection. It was found that the combination of discrimination analysis and multiple linear regression method was suitable here. Using selected membranes, the estimation accuracy of sensory quantity was raised.
