Development and evaluation of “powerful” and “refined” neural networks for the efficient choice of automotive sound quality targets

摘要

Car manufacturers want the sound of their new vehicles to appeal to potential customers. They can easily create many candidate sounds, but have to choose one target, which usually involves subjective sound quality evaluation by potential customers. Methods suitable for non-experts, like the method of paired comparisons, require considerable effort. An efficient method to predict automotive sound quality attributes of importance to manufacturers, would therefore save considerable time, money and effort. This paper describes the development of multi-layer perceptron neural networks to predict powerfulness and refinement from objective metrics of the sounds.Much effort has been devoted to select suitable treatments of time-varying metrics to use as neural network inputs. Selection methods include juror interviews, correlation, regression and empirical neural network training experiments. Novel methods have been developed to use and evaluate the neural network predictions, and to present the predicted results. Target choice is decided using win-lose graphs of pair probabilities rather than the traditional merit scores, produced by the Bradley-Terry model. The neural networks give a good insight into the relative merits of the candidate sounds. The win-lose graphs lead to better decision-making. Together, they efficiently filter candidate sounds prior to final customer testing.