Symbolic Regression for Fault Prognosis and Remaining Useful Life Estimation*

Abstract

We present a hybrid scheme for prognostics and system health management, which combines system modeling methods and regression-based approaches. Along these lines, we perform parameter trending using symbolic regression, by implementing a genetic programming algorithm that integrates the system model based on the available sensor data. The obtained fault function is an analytical expression for the progression of the system fault in time, which provides valuable insights on its causality. For comparison purposes, we also employ a dynamic degradation regression model that encompasses as health indicators inferential sensors that have been optimized by combining symbolic regression and information theory. To highlight the effectiveness of the proposed framework, both of the aforementioned approaches are applied to a dynamic model of a cross-flow plate-fin heat exchanger toward predicting fault occurrences and estimating the remaining useful life of the system, for various levels of measurement noise.