LAUSR

Thresholds are commonly used in condition monitoring and fault diagnosis of gas turbine engines as they are the criteria for health status discrimination. We discover that most thresholds should be varying according to gas turbine working condition, but many traditional fixed threshold designs didn’t consider this and therefore can potentially be replaced with dynamic ones for higher monitoring precision. In this paper, we propose an objective function that can be applied in any neural network model to learn to output a set of parameters for a distribution based on the maximum likelihood theory, and a matching dynamic threshold determination method for health state discrimination in example of a residual-based monitoring approach. The proposed methods are examined progressively with simulated data and finally an industrial gas turbine failure case. Results show that 1) the proposed objective function can learn the parameters of the distribution well; 2) the dynamic thresholds can effectively reduce the false-positive and false-negative rates when there is a varying noise comparing to fixed thresholds.