The impact of random parameter distribution on RVFL model performance in bearing fault diagnosis

Abstract

Although deep learning has achieved revolutionary progress in may application fields including image processing and fault diagnosis scenarios, the extremely high computation cost with long running time and the non-interpretability of black box mechanism seriously downgrades its applicability in some areas especially in real time tasks. To overcome these drawbacks, lightweight neural networks with quick training and less computational cost, which include the randomly weighted networks such as RVFL with non-iterative training mechanism, have been proposed to replace the deep learning models. This paper conducts a study on RVFL model performance with respect to bearing fault diagnosis problems, focusing on the difference between uniform and normal distributions of random parameters. It experimentally shows that difference of performance between both distributions is significant, and the accuracy of normal distribution is slightly better than the uniform distribution regarding this application scenario. It provides some useful guidelines for real bearing fault diagnosis with real time requirement to choose an appropriate distribution of parameters for random weighted neural network training.