Prognostics on Noisy and Uncertain Dynamic Systems using Cumulative Sum Chart of Inferential Sensors

Abstract

Abstract Inferential sensing techniques have been proven very successful in the field of fault detection and isolation, as they yield robust outcomes in a cost-effective and reliable manner. Herein, we propose a framework that deploys inferential sensors in the presence of noise and uncertainty, while focusing on the detection time and early sensitivity to faults. The most informative inferential sensor is derived through symbolic regression with an objective function that uses optimality criteria from information theory, wherein the sensitivity of the inferential sensors with respect to faults and uncertainty is estimated using the system digital twin. For deployment of the inferential sensors, the Cumulative Sum Control Chart method is employed and tuned to monitor deviations from the anticipated system performance. The proposed method is applied for the detection of faults in a crossflow plate-fin heat exchanger, at various levels of measurement noise and uncertainty, under transient operation. When compared to existing (hard) sensors, the inferential sensor provides intelligible deviations from the “fault-free” system response, thus enabling accurate and robust estimates for the initiation and progression of faults.