Echo State Network for the Remaining Useful Life Prediction of a Turbofan Engine

Marco Rigamonti, Piero Baraldi, Enrico Zio, Indranil Roychoudhury, Kai Goebel, and Scott Poll
Submission Type: 
Full Paper
phmec_16_038.pdf1.35 MBJune 27, 2016 - 1:51am

Among the various data-driven approaches used for RUL prediction, Recurrent Neural Networks (RNNs) have certain prima facie advantages over other approaches because the connections between internal nodes form directed cycles, thus creating internal states which enables the network to encapsulate dynamic temporal behavior and also to properly handle the noise affecting the collected signals. However, the application of traditional RNNs is limited by the difficulty of optimizing their numerous internal parameters and the significant computational effort associated with the training process. In this work, we explore the use of the Echo State Network (ESN), a relatively new type of Recurrent Neural Network (RNN). One of the main advantages of ESN is the training procedure, which is based on a simple linear regression. Unlike traditional RNNs, ESNs can be trained with fairly little computational effort, while still providing the generalization capability characteristic of RNNs. In this paper, we use Differential Evolution (DE) for the optimization of the ESN architecture for RUL prediction of a turbofan engine working under variable operating conditions. A procedure for pre-processing of the monitored signals and for identification of the onset of acceleration of degradation (i.e., the so-called elbow point in the degradation trend) will be shown. The datasets used to validate the approach have been taken from the NASA Ames Prognostics CoE Data Repository. These datasets were generated using a turbofan engine simulator, based on a detailed physical model that allows input variations of health-related parameters under variable operating conditions and records values from some specific sensor measurements. The results obtained on these data confirm the ESN’s capability to provide accurate RUL predictions.

Publication Year: 
Publication Volume: 
Publication Control Number: 
Page Count: 
Submission Keywords: 
data driven prognostics
echo state networks
recurrent neural networks
CMAPPS datasets
differential evolution
Submission Topic Areas: 
Data-driven methods for fault detection, diagnosis, and prognosis
Submitted by: 

follow us

PHM Society on Facebook Follow PHM Society on Twitter PHM Society on LinkedIn PHM Society RSS News Feed