Evaluation of the Training Process of three different Prognostic Approaches based on the Gaussian Process

Christian Preusche, Christoph Anger, and Uwe Klingauf
Submission Type: 
Full Paper
AttachmentSizeTimestamp
phmce_14_013.pdf321.08 KBMay 23, 2014 - 12:48am

Data-driven prognostic approaches like Gaussian Process combined with Unscented Kalman Filter (GPUKF) are promising methods for predicting the Remaining Useful Lifetime (RUL) of a degrading component. Whereas the Gaussian Process (GP) is appropriate to derive a suitable degradation model by means of a set of training data, the Unscented Kalman Filter (UKF) employs this model to determine the prediction and its uncertainty.
Since a degradation process is highly stochastic, it is assumed that by applying more sets of training data the accuracy and precision of the GPUKF is increased. In order to examine the performance enhancement two different approaches are investigated in this paper: First, a single GP is trained with all available data sets. The second approach combines several GPs (each created with a data set of one degradation process) by extending the GPUKF with a Multiple Model Method. The development of a third prognostic approach aims at the investigation of the UKF as a suitable tool for the prognostic algorithm. Therefore, a third method applies a Particle Filter in combination with the GP.
For the evaluation of the aforementioned prognosis algorithms according to their precision and accuracy a set of prevalent performance metrics like the Prognostic Horizon and the Mean Average Percentage Error of a prediction is analyzed. The validity of the determined results is increased by considering the variance of certain metrics over several units under test. Moreover, particular focus is set on the examination of the performance change caused by the use of more training data sets. In order to quantify this process known metrics are extended. The evaluation is based on simulated data sets, which are generated by an exponential degradation model.
The analysis of the implemented algorithms indicates that the applied metrics are in a comparable range. However, the three approaches reveal a different behaviour concerning the convergence of the performance values according to the number of training data. In particular cases there is even a decline in accuracy and precision attend by a rising number of training data.

Publication Year: 
2014
Publication Volume: 
5
Publication Control Number: 
013
Page Count: 
12
Submission Keywords: 
Gaussian Process Model
particle filter
unscented Kalman filter
performance metrics
performance evaluation
Prognostic Evaluation
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