Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
Spectral kurtosis-based unsupervised learning method for estimating remaining useful life of rotating machinery
Submission Author:
Leonardo Godói , SP
Co-Authors:
Leonardo Godói, Eurípedes Nóbrega
Presenter: Leonardo Godói
doi://10.26678/ABCM.COBEM2021.COB2021-0193
Abstract
Predicting the remaining useful life of rotating machines has been established as one of the biggest challenges in the area of machine health monitoring. In the last few years, a wide variety of techniques have been developed aiming at the prognosis of rotating machines, given their importance in different types of industrial processes and the subsequent criticality of failures related to specific mechanical components such as bearings and gearboxes. In this scenario, supervised and unsupervised machine learning methods have assumed great relevance due to their performance and flexibility compared to traditional ones. Recent works have studied the applicability of statistical tools for extracting features from raw signals collected from the vibration of such machines. Kurtosis, as the fourth central statistical moment, may indicate non-Gaussian tail behavior in a signal. The spectral kurtosis can find the respective components in the frequency domain that are responsible for this behavior. Therefore, it is possible to generate a spectral kurtosis-based health indicator model for the machines to be used to estimate a prognosis. The objective of this work is the proposal of a novel method for the estimation of the remaining useful life of rotating machinery through submission to an unsupervised neural network of the spectral kurtosis analysis of vibration signals, as a time-varying sequence of features, in order to identify any abnormal behavior and predict operation failures. The analysis is based on kurtogram images, which represent the spectral kurtosis calculated for multiple frequency domain window sizes along the machine operation. The images are used as inputs to a deep autoencoder neural network. The remaining useful life is estimated based on the distance between a set of samples at a given time and a normal behavior model, defined by the error resulting from the autoencoder reconstruction process. Vibration data from a public dataset containing signals extracted from faulty rolling bearings is analyzed to assess the method performance. The achieved results permit to affirm that the proposed method may be successfully adopted to estimate the remaining useful life of rotating machines.
Keywords
Spectral Kurtosis, Artificial Intelligence, neural networks, rotating machines, prognostics
DOWNLOAD PDF VIEW PRESENTATION
