A methodology for data-driven modeling and prediction of the drag losses of wet clutches

Abstract

In wet clutches, load-independent drag losses occur in the disengaged state and under differential speed due to fluid shearing. The drag torque of a wet clutch can be determined accurately and reliably by means of costly and time-consuming measurements. As an alternative, the drag losses can already be precisely calculated in the early development phase using computing-intensive CFD models. In contrast, simple analytical calculation models allow a rough but non-time-consuming estimation. Therefore, the aim of this study was to develop a methodology that can be used to build a data-driven model for the prediction of the drag losses of wet clutches with low computational effort and, at the same time, sufficient accuracy under consideration of a high number of influencing parameters. For building the model, we use supervised machine learning algorithms. The methodology covers all relevant steps, from data generation to the validated prediction model as well as its usage. The methodology comprises six main steps. In Step 1, the data is generated on a suitable test rig. In Step 2, characteristic values of each measurement are evaluated to quantify the drag loss behavior. The characteristic values serve as target values to train the model. In Step 3, the structure and quality of the dataset are analyzed and, subsequently, the model input parameters are defined. In Step 4, the relationships between the investigated influencing parameters (model input) and the characteristic values (model output) are determined. Symbolic regression and Gaussian process regression have both been proven to be suitable for this task. Lastly, the model is used in Step 5 to predict the characteristic values. Based on the predictions, the drag torque can be predicted as a function of differential speed in Step 6, using an approximation function. The model allows a user-oriented prediction of the drag torque even for a high number of parameters with low computational effort and sufficient accuracy at the same time.

Experimental investigation of drag loss behavior of dip-lubricated wet clutches for building a data-driven prediction model

“…The latest research covers experimental investigations of the integral drag loss behavior 6 , 7 , in-depth investigations 4 , 8 , 9 of the oil flow and drag torque generation in the gaps, and applying advanced technologies 10 , 11 . Additionally, research is done on analytical 12 , 13 , numerical 5 , 9 , 14 , 15 , and data-driven 5 , 16 models to calculate the drag losses.…”

“…With increasing differential speed, the drag torque first increases to a maximum value (Phase 1a), then drops to a low level (Phase 1b), remains at an almost constant value (Phase 2), and may re-increase abruptly at high speeds (Phase 3).

Figure 1 Characteristic drag torque curve of a wet clutch and its classification 16 .

…”

“…CFD models 28 – 35 and analytical models 35 – 42 are available to determine the drag losses of wet clutches. CFD models are considered highly accurate but computing-intensive, whereas analytical calculation models allow a rough but quick estimation 16 . The CFD calculation of the drag torque at a specific differential can last up to 8 h on a workstation with 32 cores 28 .…”

“…Further, the above-cited CFD and analytical models are limited to injection lubrication. In contrast, data-driven modeling enables building prediction models for any complex physical effects and lubrication type 16 . On top of that, data-driven models enable the drag losses to be predicted with low computational effort and, at the same time, sufficient accuracy 16 .…”

“…In contrast, data-driven modeling enables building prediction models for any complex physical effects and lubrication type 16 . On top of that, data-driven models enable the drag losses to be predicted with low computational effort and, at the same time, sufficient accuracy 16 . Drag torque measurements from systematic investigations on the influence of various geometry and operating parameters on drag loss behavior are needed to build the prediction models 16 .…”