Neural networks for sensor management and diagnostics

Abstract
The application of neural network technology to control system input signal management and diagnostics is explored. Control systems for critical plants, where operation must not be interrupted for safety reasons, are often configured with redundant sensing, computing, and actuating elements to provide fault tolerance and ensure the required degree of safety. In one such test case, the validation, selection and diagnosis of redundant sensor signals required 40%-50% of the control system hardware and software, while the control algorithms required less than 20% of these resources. Neural networks are investigated to determine if they can reduce the computational requirement and improve the performance of control system input signal management and diagnostics. Four neural networks are investigated to determine if they can reduce the computational requirement and improve the performance of control system input signal management and diagnostics. Four neural networks were trained to perform signal validation and selection of redundant sensors, sensor estimation from the data redundancy among dissimilar sensors, diagnostics, and estimation of unobservable control parameters. The performance of the neural networks are compared against plant models, and the results are discussed.

This publication has 2 references indexed in Scilit: