Ensemble of optimized echo state networks for remaining useful life prediction

Abstract

The use of Echo State Networks (ESNs) for the prediction of the Remaining Useful Life (RUL) of industrial components, i.e. the time left before the equipment will stop fulfilling its functions, is attractive because of their capability of handling the system dynamic behavior, the measurement noise, and the stochasticity of the degradation process. In particular, in this paper we originally resort to an ensemble of ESNs, for enhancing the performances of individual ESNs and providing also an estimation of the uncertainty affecting the RUL prediction. The main methodological novelties in our use of ESNs for RUL prediction are: i) the use of the individual ESN memory capacity within the dynamic procedure for aggregating of the ESNs outcomes; ii) the use of an additional ESN for estimating the RUL uncertainty, within the Mean Variance Estimation (MVE) approach. With these novelties, the developed approach outperforms a static ensemble and a standard MVE approach for uncertainty estimation in tests performed on a synthetic and two industrial datasets.

An ensemble of models for integrating dependent sources of information for the prognosis of the remaining useful life of Proton Exchange Membrane Fuel Cells

“…Fusing the outputs of an ensemble of diverse prognostic models can improve overall prediction accuracy [34]. Local aggregation dynamically assigns weights to each model according to its local performance, typically evaluated on the available historical patterns [25]. For prognostics, local aggregation requires the computation of the local performances of the individual models on a set of run-to-failure degradation trajectories.…”

“…In this work, we develop a hybrid approach based on an ensemble of models, which uses prediction outcomes provided by different degradation models fed by different deterioration measurements and properly combines them to provide the prognostic results. Ensemble of models have shown promising results for the prognostics of industrial systems [20,21,22,23,24,25]. For example in [23], an ensemble approach based on a semi-Markov model and a fuzzy similarity model has been developed for the predictions of the RUL of a heterogeneous fleet of aluminum electrolytic capacitors used in electric vehicle power trains.…”

An evidential similarity-based regression method for the prediction of equipment remaining useful life in presence of incomplete degradation trajectories

“…We consider simulated degradation trajectories of a heterogeneous fleet of N = 218 turbofan engines operating under variable operational conditions. The data are taken from the NASA Ames Prognostics CoE Data Repository [39] and have been preprocessed as in [15], obtaining M = 6 relevant features describing the component degradation evolution. Fig.…”

“…To deal with this type of data, we have applied a pre-processing procedure followed by a feature extraction step, which has allowed the identification of M = 2 relevant features describing the component degradation evolution. More details on the dataset and feature extraction procedure can be found in [15,40]. Fig.…”

“…When few degradation trajectories with no clear patterns of regularity are available for model training, similarity-based (also known as instance-based) learning algorithms [13] have been proved to be effective given that they do not perform explicit generalization. Contrarily, non-linear regression methods such as those based on Artificial Neural Networks (ANN) [14,15], which are characterized by large numbers of parameters to be tuned, tend to underperform due to the scarcity of training examples [16]. Indeed, direct RUL prediction approaches assume the availability of a set of complete run-to-failure degradation trajectories for training.…”