On-Line Learning of Evolving Neural Models for Process Identification and Abnormality Detection

In this paper a new algorithm for creating evolving neural models is proposed. Instead of repeating the "growing" and "pruning" steps during the learning, as in the most other known evolving algorithms, here we create a growing neural model from a fixed size data buffer and repeatedly check the model quality, in the sense of "average minimal distance" between the neurons and the data in this buffer. If the current neural model does not satisfy the required quality for the new on-line updated data buffer, then a new growing neural model is created from the updated data buffer. The new proposed algorithm is very efficient in computation cost. Its better performance is demonstrated by comparison with two other algorithms with preliminary fixed number of neurons, namely the standard non-evolving learning and the real-time evolving algorithm. The results show the applicability of the new algorithm for modeling and detection of changes and abnormalities in real-time evolving processes.

1. Introduction

The problem of process identification and modeling is encountered in many applied areas such as control, pattern recognition, communications and performance evaluation for maintenance and fault diagnosis.