Frequency Domain Signal-to-Noise Ratio Pre-Processing for Fault Classification in Planetary Gear Systems Using Audio Data

Rotating machines commonly used in industries, such as planetary gears transmission systems, emit acoustic signals that contain significant information relating to the equipment's condition. Thus, the acquisition, processing and analysis of these signals are interesting procedures to enable the machine's working state monitoring. Given this scenario, we propose a technique for fault classification of planetary gear systems, based on a Frequency Domain Signal-to-Noise Ratio (FDSNR) pre-processing algorithm, applied over the audio signals emitted from the gear system. The processing procedure feeds Supervised Machine Learning models such as Artificial Neural Networks and XGBoost. For our practical experiment, a planetary gear system was built in the laboratory, using a set of 3D-printed plastic gears and physical components and structures used for fixation and assembling purposes. By printing some anomalous gears, we could make replacements to alternate ordinary gears with anomalous ones, providing three different conditions for the apparatus: one normal configuration, and two anomalous conditions induced by a mass imbalance caused by the defective gears. Our final approach used: a microphone and an audio interface for acoustic capturing and audio recording; acoustic processing techniques that extract the key features from the audio files; and machine learning models that use these features to perform fault classification. The final experiment using the approach achieved a peak accuracy of 98% with an XGBoost model.