A neural network approach to the automated detection of acoustic annoyance in hard disk drives

Abstract

In order to increase efficiency, computing equipment in the office environment is increasing rapidly. However, the increase in efficiency is not optimized since the added equipment also brings added noise to the environment. "Annoyance" in this research is defined as the reduction of productivity due to the acoustic disturbance in the office environment. This paper describes an artificial neural network based approach for automating the detection of annoyance in hard disk drives. The aim is to be able to detect excessive annoyance as a quality test while still in the manufacturing environment. The true question of this research is how to map subjective information to a quantitative domain. First, the hard disk drive sounds were analyzed and characterized using both time and frequency domain analysis. Literature suggested that magnitude, time variance, and frequency combinations were all contributors to the effect of annoyance; however, a deterministic method could not be found to predict the level of annoyance from this information. The hard drives were divided into two groups of annoying and not annoying. The Learning Vector Quantization and the Back Propagation algorithms were both used to train networks to perform this decision function. A discussion is presented comparing the two methods and their training results.