Using Neural Networks to Detect Failure Onset in Complex Systems

Successful prognostic and health monitoring systems depend on being able to recognize the signs of a failure in progress. Although such systems are commonplace, little has been reported to date on fault detection for systems where the interactions of the various operating parameters are subtle, complex, and correlated in unknown or difficult to elicit ways. This paper describes the results of recent research into the use of neural networks to provide detection of the onset of operational failure in such devices. After a preliminary exploration revealed the shortcomings of more common pattern recognition methods, such as limit checking, a posteriori Bayesian methods, and even principal component analysis, it is shown that certain types of neural networks are up to the task. The results from simulations will show the effectiveness neural network techniques in detecting the onset of the failure. These techniques will then be demonstrated on data from a real-world system and the results presented.