Dear Sina,

You are correct to note that dynamic instability problems like NVH can be notoriously difficult to model because of the complex physics and presence of many variables. You are also right to note that traditional techniques to study NVH, like chassis dynamometer tests and computational modelling, are very difficult to implement and time consuming.

One of the most valuable things machine learning does is feature reduction; it can reduce a very complex problem into a manageable set of parameters. There is a ton of work to be done in this field for NVH, but some researchers have started, and the future outlook is promising. Machine learning can certainly be accurate for NVH if the model is designed well and features are neither over or under trained.

Please take a look at this thesis that provides a very good overview of current progress:

https://deepblue.lib.umich.edu/handle/2027.42/163518

Hope this helps!

Conventional methods like computational techniques are useful for NVH analysis. ML is new tool for NVH analysis and researchers are working in this emerging area.

Please refer the following detailed thesis for use of ML in NVH analysis.

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

https://hdl.handle.net/2027.42/163518

Dear Dina,

Please find the attached article for your reference.

Hope it will be useful.

Thank you

Nathan Zavanelli Ashwani Kumar I am not still convinced with the efficiency of ML for NVH prediction. Since I deal with optimal tuning of NVH attenuators, advent of local modes or separation of modes with a specific distance at not-studied points is probable. This is what cannot be predicted by ML, but only by an eigenvalue analysis.

Nathan Zavanelli I understand your statement "...if the model is designed well"... but the main issue is that in NVH problems specifically in real applications a wide band of frequency is studied. Then, if someone is going to train a system for a frequency range of 0 to 5k Hz, and the ML should be trained for half of this frequency range, there is not difference in computation cost saving compared to the conventional analysis.

I think ML be used for super limited frequencies ranges if a reasonable computation cost saving is expected.

Thanks Nathan & Zavanelli for your attention.