Data-Driven Soft Sensor for the Estimation of Sound Power Levels of Refrigeration Compressors Through Vibration Measurements

Abstract

The quality assessment of the sound power generated by refrigeration compressors is an important and costly process for manufacturers, mainly due to the need for acoustic rooms. Hence, very few samples per lot are tested. In this article, an alternative method based on soft sensing is proposed to estimate the sound power of refrigeration compressors. The method uses vibration measurements from the compressor surface as the input and the sound power generated as the output of machine learning (ML) models to create data-driven soft sensors. This method is less restrictive than the traditional ones and allows a test to be done close to the manufacturing line, in a much faster and cheaper way. In order to create the dataset to train the ML models, sound power and vibration levels of compressors were measured simultaneously in a reverberation room. The results from the data-driven soft sensors were compared with the ones obtained using a simple analytic model defined in an ISO technical note and showed that, even with as little as one vibration measurement point, the proposed method presented better results than the ISO analytic model with nine measurement points.