RULE EXTRACTION FROM A KOHONEN SELF-ORGANISING MAP FOR EQUIPMENT CONDITION ASSESSMENT USING NOISY DIAGNOSTIC SIGNALS

Abstract

This paper proposes a method for extracting classification rules for one-dimensional (1D) diagnostic signals from a trained Kohonen self-organising map (SOM). To validate equipment condition assessment from time-series data, a classification problem involving second-order curves with similar fragments was formulated. A balanced dataset was generated from a mathematical model, the SOM was trained, and the corresponding clusters were identified using a multilayer perceptron (MLP). Rules were extracted using a decompositional approach in IF–THEN form, where conditions were defined in terms of the best-matching units (BMUs) of the input signal. The proposed approach enables robust classification of noisy signals. A software suite for populating a knowledge base with extracted rules was developed. Computational experiments on signal classification were conducted using both the SOM and a CLIPS-based rule-based system, with the number of antecedent conditions and the noise factor serving as simulation parameters. The results show that, with the maximum number of antecedent conditions, the classification accuracy of the rule-based system decreases by 1–3 percentage points depending on the noise factor.