Prognostic Information Fusion for Constant Load Systems Kai Goebel GE Global Research K1-5C4A, One Research Circle, Niskayuna, NY 12309 USA goebelk@research.ge.com Piero Bonissone GE Global Research K1-5C32A, One Research Circle, Niskayuna, NY 12309 USA bonissone@research.ge.com Abstract –This paper describes a process for aggregating different information sources to estimate remaining equipment life. Specifically, the approach presents a rigorous chain of preprocessing, modeling and post- processing steps that arrive at the desired prognostic result. The preprocessing steps deal with data reduction, filtering, and signature amplification. The prediction model applies Adaptive Neuro-Fuzzy Inference System (ANFIS) to the data. The post-processing steps include recursive trending which implicitly forces the prognostic trend to be confirmed before updated estimates are reported. Prognostic false positives and false negatives are introduced as innovative measures that help in assessing the performance of the approach. The method is illustrated using real-life data from industrial web paper breakage prediction. Keywords: Prognostic fusion, prognostics, prognosis, decision fusion 1 Introduction Finding synergy in using different information sources to assess system states has a long tradition within the fields of multivariate statistics and pattern recognition. Recently, the field of information fusion, and more specifically multi- classifier fusion has been recognized as a research area in its own right. Fusing information for prognostic purposes is a fairly new endeavor and will likely lead to the development of new techniques that are specialized to perform related tasks. Prognosis is a key element in equipment health management by providing an estimate for remaining equipment life. Accurate prognosis is difficult to achieve for a number of reasons. These range from the availability of suitable data sets for training of algorithms to the unsolved issue of validating prognostic approaches to the incorporation of future usage information to the need to manage uncertainty to the need to find technology that aggregates diverse information sources and returns a continuous output reflective of remaining life. This paper will address some aspects of the latter issue. Generally, remaining life estimates are conditional on future usage. That is, the estimate must change for real-time estimates as new information about intended usage becomes available. For many systems, future usage is less certain the farther out the prediction is. This means that remaining life will also have a much wider range of possible values and the uncertainty about any specific estimate increases. The situation is slightly better for constant load processes where few process changes are carried out or where the changes are largely the same. The work reported herein – prediction of the next failure at the wet end of a paper mill – faced the challenge that none of the dozens of observable process and control variables provided a single strong indication for an impending failure. Indeed, there seemed to be not even an apparent consistent weak indicator. Some variables looked as if they could provide a trend for particular failure trajectories, but that was not true for many other failure trajectories. While this may be due to different failure root causes, it might also point to a lack of single variable predictions. With a lack of access to further domain knowledge, the problem could be posed as a fusion problem where the task was how to fuse dozens of process variables to arrive at a reliable failure prediction. 2 Prognostics Using the definition where prognostics is the estimation of remaining useful component life, the task of providing that estimate can be accomplished using three fundamental different means. 1. The first is extrapolation from past data. Most statistical approaches fall into this category. These include • Linear extrapolation • Non-linear curve fit (extrapolation) • Multi-variable regression • Auto-Regressive Moving Average (ARMA, ARIMA) Models • Multi-Step Adaptive Kalman Filter (Frelicot, 1996) • Variance Analysis (Fuh and Wu, 1995) • Particle Filter (Bauer at al., 2002) • Least angle regression (LARS) (Efron et al., 2004) • Shrinkage methods (Ojelund et al., 2002 ) • Least Absolute Shrinkage and Selection Operator (LASSO). (Tibshirani, 1996). 0-7803-9286-8/05/$20.00 © 2005 IEEE