A non-tachometer method for Order Tracking technique in NVH analysis based on Deep Learning and rpm estimation

Abstract

This paper presents a new analysis method of automobile noise-vibration-harshness(NVH) analysis based on a discrete recurrent neural network(RNN) and generative adversarial network(GAN), which, can not only replace Short Fast Fourier Transform(SFFT) but also the entire tachometer data assembly system for our network's ability to obtain rpm from vibration signal. This method inherits the leading spirit of digital resampling and Time-Variant Discrete Fourier Transform(TVDFT), adjusting sampling rate concerning rpm changes and interpolation to obtain an equal time interval sequence out of identical angle interval sequence, as the setting parameter of these methods determines the quality of order tracking. The neural-network-based approach involves three steps: 1. Simulation and sampling of the vibration signal of a DeLaval rotor. 2. Determination of rpm, and instantaneous sampling rate, window size as well as resampling time and values through a discrete RNN-GAN learning system with the input vibration signal and output parameters. 3. Illustration of a dB-rpm graph obtained by D-RNN-GAN and further evaluation of system performance 4. The application to big data and its review.