Prediction of Failure in Lubricated Surfaces Using Acoustic Time–Frequency Features and Random Forest Algorithm

Scuffing is one of the most problematic failure mechanisms in lubricated mechanical components. It is a sudden and almost not predictable failure that often leads to extensive cost in terms of damages and/or delay in production lines. This study presents a promising solution that can prevent scuffing for the machinery industry in the future. To achieve this goal, a signal processing approach by means of an acoustic emission is introduced for the prediction of scuffing. An acoustic dataset was collected from metallic surfaces reciprocating under a constant load (typical conditions for semi journal bearings). The coefficient of friction values were measured during the entire experiments and were referred to as the ground truth of the momentary surface state. Based on the friction behavior, three friction regimes were defined that are running-in, steady-state, and scuffing. The present approach is based on tracking the changes in acoustic emission by means of three sets of wavelet-derived features. Those features include: 1) energy, 2) entropy, and 3) statistical information about the content of acoustic emission and the response of each feature to the different friction regimes was individually investigated. The applicability of machine learning classification and regression was studied for scuffing prediction. Both approaches were applied separately but can be unified together to increase the prediction time interval of surface failure. For classification, an extra friction regime was introduced designating as pre-scuffing and is defined as a time span of 3 min before the real surface failure. Random forest classifier was used to differentiate the features from the different friction regime. The best performance in classification of features from pre-scuffing regime reached a confidence level as high as 84%. In regression approach, the extracted features sequences were used together with random forest regressor. Our strategy allowed predicting scuffing up to 5 min preceding its real occurrence.