Implementation of Machine Learning in Acoustics Source Detection by Leveraging Synthetic Sound Data Generation Approach

Abstract

<div class="section abstract"><div class="htmlview paragraph">E-Mobility and low noise IC Engines has pushed product development teams to focus more on sound quality rather than just on reduced noise levels and legislative needs. Furthermore, qualification of products from a sound quality perspective from an end of line testing requirement is also a major challenge. End of line (EOL) NVH testing is key evaluation criteria for product quality with respect to NVH and warranty. Currently for subsystem or component level evaluation, subjective assessment of the components is done by a person to segregate OK and NOK components. As human factor is included, the process becomes very subjective and time consuming. Components with different acceptance criteria will be present and it’s difficult to point out the root cause for NOK components. In this paper, implementation of machine learning is done for acoustic source detection at end of line testing. To improve the fault detection an automated intelligent tool has been developed for subjective to objective method conversion using ML model to segregate OK and NOK components and its root causes. However, the key challenge with ML models is its reliance on significant amounts of data that are subjectively tagged. This paper talks about a unique approach to generate large synthetic sound data from smaller set of tagged sound data sets.</div></div>