Convolutional and long short-time memory network configuration to predict the remaining useful life of rotating machinery

Abstract

Recently, several machine learning approaches have been proposed to provide predictions of the remaining useful life of rotating machine. This study presents a strong framework that employs machine learning algorithms to predict the useful life of rotating machine bearings by evaluating their vibration signals. In this approach, the raw vibration signal undergoes feature extraction through auxiliary methods, trend analysis through statistical methods, and time-dependent feature extraction through a specialized hybrid neural network algorithm. The architecture is composed of three distinct phases: Feature analysis, where the raw vibration data are processed to extract important characteristics for the definition of the signal trend creating a time series and Modeling, where the training data is processed in a hybrid convolutional neural network, which returns a degradation model aiming at estimating the instant of total failure. The neural network is also utilized to analyze test data and identify the moment just prior to the occurrence of failure; and finally the Prediction, phase where the future failure trend of the test data is identified, using the failure threshold extracted from the training data. We used the architecture to predict the remaining useful life of rotating machines in various cases, and the results error ranged between 3 and 4%, which is considered a good result.