Real Time Knowledge Acquisition Based on Unsupervised Learning of Evolving Neural Models

Abstract

This paper presents a method for extraction of knowledge from a real time process by using the so called evolving neural model (ENM). The ENM learns from real time data streams by a specially proposed evolving unsupervised learning algorithm. This algorithm is further development of the off-line neural-gas learning with a different way of updating the neurons. It also uses a special logic to prevent the neurons from gradually becoming "idling" during the evolutions. Two characteristics of the ENM, namely the center-of-gravity COG and the weighted average size WAS of the model are further used to capture the general trends of operation changes in the process. Big changes serve as indication for acquisition of a new knowledge about the process that should be saved in the knowledge base. Normalized data taken from different operations of a diesel engine for hydraulic excavator are used to test and verify the merits of the proposed learning algorithm and the whole knowledge acquisition method.