Implementation of Machine Learning Techniques in Brake NVH Using Computational and Experimental Data

A complex system contains a broad range of intrinsic properties and parameters. Involved interconnections between parameters make it extremely difficult to decompose the system into pieces and predict its behavior. In automotive NVH, brake squeal is a dynamic instability phenomenon that contains complex physics due to its nature of the problem and complicated interconnections and variations; Also includes a wide range of operational conditions over a broad range of frequencies. Squeal noise is one of the most significant customer claims results in high warranty costs for vehicle manufacturers. Engineers study brake NVH using city and chassis dynamometer tests and computational techniques. These methods are costly and time-intensive. In recent years, by advancements in CAE tools and resources, numerical methods are vastly employed; They can replace physical experiments. This research proposes and implements an innovative Machine Learning based technique on brake noise analysis and accelerates computational methods. The proposed approach introduces a new metric that explores correlations of operating condition distributions from the virtual and physical models. Data-driven methods incorporated with numerical simulations have been increasingly developed and utilized in recent years to improve virtual models' efficiency. This work demonstrates an ML-based model can significantly save computational cost by exploring an entire design space and skipping over duplicative iterations. The second part of the research proposes and implements a multi-fidelity Deep Learning approach to predict brake pad NVH modal characteristics. This approach, inspired by both physics and statistics. It leads to a perception of component properties, and sequentially, their modal responses. This work initially develops a high-fidelity numerical model and correlates with experimental data capturing brake pad component physical aspects. Then uses a Design of Experiment technique to generate a high-resolution database from CAE. This database is ultimately used to develop a physics-inspired deep learning model. The deep learning algorithm is composed of distinct multi-layer perceptron (MLP) modules that define component properties. The goal of this ML-based approach is to accelerate design and development processes in brake NVH and minimize urgency to additional experiments and simulations. This methodology may apply to other components and full system analysis.PhDCollege of Engineering & Computer ScienceUniversity of Michigan-Dearbornhttp://deepblue.lib.umich.edu/bitstream/2027.42/163518/1/Morteza Gorzinmataee Final Thesis.pdfDescription of Morteza Gorzinmataee Final Thesis.pdf : Dissertatio