A probabilistic, ``neural'' approach to sensor modelling and classification is described, performing local data fusion in a wireless system for embedded sensors using a continuous restricted Boltzmann machine (CRBM). The sensor data clusters are non-Gaussian and their classification is non-linear. A CRBM is shown to be able to model complex data distributions and to adjust autonomously to measured sensor drift. Performance is compared with that of single layer and multilayer neural classifiers. It is shown that a CRBM can resolve the problem of catastrophic interference that is typical of associative memory based models. r 2007 Elsevier B.V. All rights reserved.
Tong Boon Tang, Alan F. Murray