Condition classification of heating systems valves based on acoustic features and machine learning

Abstract

The quality and condition of valves installed in district heating systems can be reflected by the sounds emitted. In this paper, a framework for a systematic approach towards the classification of valve sounds is proposed, based on acoustic features and machine learning models. The methods include the extraction of spectral and psychoacoustic features, alongside the application of a wrapper-based feature selection method which, when combined with machine learning models, simultaneously selects the most informative features and builds optimal classification models. The maximal balanced classification rate (BCR) was used as the optimisation criterion in this study. Results demonstrate that the specific valve conditions can be correctly classified with a high BCR as follows: cavitation BCR = 1, whistling BCR = 0.978, and rattling BCR = 1. The proposed framework for a wrapper-based selection of informative features and corresponding machine learning models confirms the usefulness of psychoacoustic features and machine learning models for the classification of valve conditions. The proposed framework is, however, general and can be applied to various acoustic-based industrial condition monitoring challenges.