Predicting the coefficient of friction in a sliding contact by applying machine learning to acoustic emission data

Abstract

It is increasingly important to monitor sliding interfaces within machines, since this is where both energy is lost, and failures occur. Acoustic emission (AE) techniques offer a way to monitor contacts remotely without requiring transparent or electrically conductive materials. However, acoustic data from sliding contacts is notoriously complex and difficult to interpret. Herein, we simultaneously measure coefficient of friction (with a conventional force transducer) and acoustic emission (with a piezoelectric sensor and high acquisition rate digitizer) produced by a steel–steel rubbing contact. Acquired data is then used to train machine learning (ML) algorithms (e.g., Gaussian process regression (GPR) and support vector machine (SVM)) to correlated acoustic emission with friction. ML training requires the dense AE data to first be reduced in size and a range of processing techniques are assessed for this (e.g., down-sampling, averaging, fast Fourier transforms (FFTs), histograms). Next, fresh, unseen AE data is given to the trained model and the resulting friction predictions are compared with the directly measured friction. There is excellent agreement between the measured and predicted friction when the GPR model is used on AE histogram data, with root mean square (RMS) errors as low as 0.03 and Pearson correlation coefficients reaching 0.8. Moreover, predictions remain accurate despite changes in test conditions such as normal load, reciprocating frequency, and stroke length. This paves the way for remote, acoustic measurements of friction in inaccessible locations within machinery to increase mechanical efficiency and avoid costly failure/needless maintenance.