(614f) Fault Detection and Isolation in Uncertain Dynamic Systems Using Sensor Fusion and Inferential Sensing

The effective operation of modern cyber-physical systems is linked to increased requirements, pertaining to safety, dependability and lower cost. To satisfy safety and operability requirements, built-in test (BIT) design methods for fault detection and isolation (FDI) are often employed. The increasing complexity and nonlinearity of modern systems, however, as well as the presence of noise and uncertainty in data, effectuate a need for more effective approaches that yield robust fault identification and isolation during system operation. In this regard, inferential sensing approaches are cost-effective and reliable alternatives to expensive and often impractical measuring devices, as they use information already available from other measured variables and system parameters to estimate a quantity of interest. In principle, inferential sensors combine available system inputs and outputs in either analytical expressions, which are founded on physical laws and knowledge of the system structure, or empirical relationships based on data, such as regression models, Support Vector Machines, Neural Networks, and genetic programs. Building upon concepts described in Hale and Bollas (2020), in this work we attempt to combine advanced model-based sensor selection and inferential sensing techniques to provide accurate fault detection, in the presence of system noise and uncertainty. During sensor selection, criteria from information theory are employed to maximize the estimability of system faults from available system measurements. The most informative sensor set is chosen, by comparing all possible sensor combinations based on various optimality criteria, which extract the maximum knowledge from the system in functions of the Fisher Information Matrix. Subsequently, optimal inferential sensors are derived, by means of genetic and mathematical programming, which are based on symbolic regression and optimization techniques. The augmented system of composite sensors (i.e., inferential and hardware) is shown to provide more accurate information about the system fault(s) and reduce the evidence of uncertainty and system noise. For BIT deployment, k-Nearest Neighbors (k-NN) classification is deployed to: (i) assess the accuracy of each sensor network for all plausible instantiations of uncertainty and system noise, and (ii) assess the performance enhancement due to the inclusion of inferential sensor(s). The proposed methods are applied in steady state and dynamic models of a cross-flow plate-fin heat exchanger system (Palmer et al. 2016) for various levels of measurement noise and uncertainty. When compared to traditional sensor approaches, the techniques developed herein prove to be more accurate, while simultaneously granting increased insight to details otherwise hidden by noise. Consequently, they constitute a robust and valuable tool when pursuing to meticulously conduct fault diagnosis and prognosis, toward system safety.

Acknowledgements

This study was supported by the UTC Institute for Advanced Systems Engineering (UTC-IASE) at the University of Connecticut (UConn).

References

Hale W.T., and Bollas G.M. (2020) Least-Squares- and Information-Theory-Based Inferential Sensor Design for Fault Diagnostics, 2020 American Control Conference (ACC), Denver, CO, USA, pp. 3182-3187, doi: 10.23919/ACC45564.2020.9147305.

Palmer K.A., Hale W.T., Such K.D., Shea B.R., Bollas G.M. (2016) Optimal design of tests for heat exchanger fouling identification, Applied Thermal Engineering, 95:382-93,