Using Gaussian process regression for building a data-driven drag loss model of wet clutches

Abstract

Drag loss calculation models support designing low-loss wet clutch systems and determining drivetrain efficiency. Available numerical models require high computational effort, whereas analytical models provide only limited accuracy. In contrast, data-driven models unite the advantages of fast and effective drag loss prediction and, thus, enable practice-oriented development. Therefore, this study aimed to build a data-driven model of the wet clutches’ drag losses, enabling predictions with low computational effort and, at the same time, sufficient accuracy while taking seven relevant influencing parameters into account. The dataset used originates from prior research projects that experimentally investigated the wet clutches’ drag loss behavior under consideration of various influencing parameters. Different machine learning algorithms were considered for model building, and it was found that Gaussian process regression best met the requirements of flexibility and interpretability. Four sub-models were built to predict four characteristic drag loss values based on the model input parameters of mean diameter, diameter ratio, number of gaps, clearance, dynamic viscosity, oil level, and groove design. Subsequently, the predicted characteristic drag loss values can be used as support points for approximating the drag loss curve. The sub-models show high R² values, ranging from 0.84 to 0.95, indicating a high model performance. The mean relative error of the sub-models was used for evaluating the models’ performance from an engineering perspective, ranging from 4.6 % to 14.2 %. The drag loss model was also validated based on documented knowledge of wet clutches’ drag loss behavior. The model enables the prediction of the drag loss curve within a few split-seconds and, simultaneously, with sufficient accuracy.