Research on super-fitting phenomenon in sound quality prediction of silent chain transmission system

Abstract

To predict the sound quality of silent chain transmission system, we propose a new data enhancement method called fuzzy generation. After enlarging the original dataset by fuzzy generation, we find that prediction model has an abnormal super-fitting phenomenon under certain conditions. In contrast to overfitting, super-fitting means that the model performs poorly on the training set but performs better on the test set. To further explore this super-fitting phenomenon, we firstly carry out a noise collection test of the silent chain transmission system. Secondly, the collected noise samples are evaluated subjectively and objectively. Based on the subjective evaluation results, we can establish the fuzzy generation interval for each sample. By selecting different membership values and different perturbation forms, the dataset is expanded three times. Training four different prediction models on the new datasets, we found that super-fitting is more likely to occur with linear perturbation. However, in the case of some membership values, the random perturbation can also lead to super-fitting. Through analyzing the prediction effect of the four models, we conclude that fuzzy generation can indeed significantly improve accuracy of model, and super-fitting is mainly caused by a special kind of data leakage. When using fuzzy generation to predict sound quality with small samples, appropriate membership value and perturbation form should be selected to avoid the occurrence of super-fitting.