Complex eigenvalue analysis (CEA) is a conventional evaluation tool in virtual brake system development for improved noise, vibration and harshness (NVH) characteristics. The discretization of full-scale brake systems, using a quarter-car model, results in millions of finite elements for which the linearized dynamic vibration analysis of CEA is carried out. Hence, the calculation of all eigenvalues across a wide frequency range for a sample space of possible operational parameters during a brake procedure regarding brake pressure, rotational velocity and coefficient of friction is compute-intense and time-consuming. This study presents a machine learning approach to compute the complex eigenvalues between 1 kHz and 5 kHz for a FE model of a disc-pad setup for a given load configuration in terms of disc rotation velocity, brake pressure and coefficient of friction. Training data are generated through a design-of-experiments study using latin hypercube sampling and a classical CEA solver. Using XGBoost machine learning approaches, we illustrate that mode-coupling instabilities can be represented successfully by the ML model for new and unseen load parameters. The near-real-time capabilities of the ML models are shown to pay off against the offline costs for training data generation, hence rendering the ML approaches particularly relevant for repetitive computations.

Volkswagen AG: Marcel Deutzer; TU Hamburg: Dr.-Ing. Merten Stender, Prof. Dr. Norbert Hoffmann