Dynamic Gas Turbine Condition Monitoring Scheme With Multi-Part Neural Network

Abstract

Gas turbines are high value industrial assets with significant roles in various kinds of industrial processes, health management systems are therefore important for helping maintaining gas turbines’ stability in long-term operations. With more and more performance data able to be collected by sensors and the new machine learning methods developed, researchers are able to build more powerful digital models to monitor the gas turbine. This paper introduces a performance parameters alarm scheme for gas turbine using an adaptive state following model. The proposed scheme consist of 3 parts: Part 1, a dynamically adaptive multi-part neural network trained using performance data that can simulate different parts of gas turbine and output “normal” sensor data to make comparison with the actual data collected; Part 2, a group of thresholds set according to system noise that flags sudden failures by sensing performance parameter outliers, this also decides which data should be used to update the neural network; Part 3, a recorder for “reference point” outputs that can reflect change of the gas turbine’s status and detect long-term degradation. Unlike traditional approaches, the proposed adaptive states following model separates long term degradation and short term sudden failure, therefore both faults can be detected more accurately. The core of the proposed method is that physical properties are embedded into the neural network as constraints to regulate training and make the model more interpretable. In our scheme, a gas turbine is divided into 4 parts referencing the equipment’s physical mechanism, they are simulated digitally by 4 sub corresponding networks, which are then combined into the proposed integrated network. The proposed scheme achieves an overall pleasing result and shows potential in gas turbine fault analysis.