Dynamic Weighted PSVR-based Ensembles for Prediction of Nuclear Components

Jie Liu, Valeria Vitelli, Redouane Seraoui, and Enrico Zio
Submission Type: 
Full Paper
AttachmentSizeTimestamp
phmce_14_009.pdf462.82 KBMay 19, 2014 - 6:53am

Data-driven approaches are widely used for prediction. They do not make use of any explicit model and rely exclusively on data measured by sensors related to the degradation and failure state of the component. The advantage of these methods lies in the direct use of the measured data for empirical learning and non-parametric estimation. However, data-driven approaches have some drawbacks limiting their applications. In the dynamic, context-changing environment of component operation, single algorithm model of this kind cannot always reconstruct the true underlying process related to the data and may cause over-fitting on the data set at hand. In addition, many machine learning algorithms are based on some form of local search that may converge to local optima. Combining approximate algorithms into an ensemble can improve the robustness and accuracy of the prediction. An ensemble is obtained by training diverse sub-models and, then, combining their results. The objective is to take advantage of each sub-model, by fusing results from all the sub-models. In this paper, an ensemble approach is proposed for prediction of time series data based on a modified Probabilistic Support Vector Regression (PSVR) algorithm. We propose a modified Radial Basis Function (RBF) as kernel function to tackle time series data set and two strategies to build diverse sub-models of the ensemble. A main advantage of PSVR is that it gives not only predicted values, but also the associated error bar estimations, by assuming a Gaussian distribution of the predicted value. A simple but effective strategy is used to combine the results from sub-models built with PSVR, giving the prediction results of an ensemble. A real case study on a power production component is presented.

Publication Year: 
2014
Publication Volume: 
5
Publication Control Number: 
009
Page Count: 
9
Submission Keywords: 
ensemble methods
Data-driven prognostics
dymanic weighting
support vector regression
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