Prediction of individual thermal sensation using unspecified sensors in sensor networks

This paper describes a prediction method for predicting human thermal sensation by using unspecified sensors over an unstable sensor network. First, a dynamically changing neural network is utilized for predicting the thermal sensation. The neural network is formed dynamically and trained by considering the strength of the correlation between the sensor readings and the thermal sensations of subjects. The neural network is modified when the difference between its estimation and the actual values increases. Next, in order to perform experiments with actual subjects, we built a sensor network in an indoor environment. For two weeks, we regularly measured certain values, such as the temperature in the environment, and investigated the thermal sensation of the subjects once every fifteen minutes while they were in this environment. Then, using our method, the thermal sensation and the thermal values were associated with each other, after which a dynamical neural network which estimates each thermal sensation was built automatically.