Trans-layer model learning: A hierarchical modeling strategy for real-time reliability evaluation of complex systems

Techniques addressing the loading condition of components in complex systems are of great significance for the real-time reliability analyses of systems. To recover component observabilities with combined condition monitoring data and empirical rules, an information criterion identifying the necessary data/rule set for the modeling of systems with the same hierarchical topologies to real-in-world realizations, referred to as trans-layer model learning (TLML), is proposed and proved. Then, with regard to general multi-component dynamic systems, a specific TLML algorithm is proposed. In this algorithm, the loss function and alternative training scheme of component models are specified for harnessing the information from sensor readings and empirical rules to serve the modeling. TLML is applied first on a simulation system to testify its ability to reveal component loading conditions, and then on an aircraft engine to test its effectiveness in improving the Residual Useful Life (RUL) prediction performance of engine turbine blades. Results show that TLML can provide real-time estimations of component loading conditions with sufficient accuracy, and thus improve the precision and reliability of the RUL estimation of system parts.