Symbolic regression of uncertainty-resilient inferential sensors for fault diagnostics

Abstract

An algorithm is presented for the design of inferential sensors for fault diagnostics in thermal management systems. The algorithm uses input and output sensed system information to improve the detection and isolation of a fault by generating inferential sensors that augment the measured information to: (i) reduce the evidence of uncertainty in the inferred variables, and thus decrease false alarm and nondetection rates; and (ii) provide distinguishable responses to faults, and thus reduce reduce the rate of misdiagnoses. The novelty of the algorithm is its use of genetic programming to evolve explainable inferential sensors that maximize information criteria specific to fault diagnostics. The chosen criteria: (i) least squares regression; and (ii) Ds -optimality (calculated from the Fisher Information Matrix), leverage symbolic mathematics and automatic differentiation to obtain parametric sensitivities of the measured outputs and inferential sensors. The algorithm is included in a standard work for fault diagnostics, where its effectiveness is assessed through k-NN classification and illustrated in an application to an aircraft cross-flow plate-fin heat exchanger.