Adaptive machine learning with physics-based simulations for mean time to failure prediction of engineering systems

•Developed a new approach for mean time to failure (MTTF) analysis of nonrepairable systems.•The developed approach employed adaptive surrogate models using Gaussian processes.•It used data from physics-based simulations to train the surrogate models.•A novel composite expected feasibility function was created as criteria for model refinement.•Case studies were used to demonstrate the developed approach in system MTTF analysis. The Mean Time to Failure (MTTF) is a critical metric for assessing the reliability of non-repairable systems, and it plays a significant role in incident management. However, accurately estimating MTTF can be challenging due to the expensive physics-based simulation models. To address this challenge, this paper proposes an adaptive surrogate modeling method that approximates the failure modes in simulation model with a computationally efficient model to predict the MTTF during the design stage. Firstly, the proposed method initially trains Gaussian process (GP) surrogate models for the failure modes. Then, the composite expected feasibility function is proposed to identify the new information, such as input variables, time instances, and component index, to refine the surrogate models. In the end, the MTTF can be calculated by taking the expected value of the system’s first time to failure with the available GP models. The proposed method has the capability of forecasting MTTF for series systems, parallel systems, and mixed systems. To showcase its efficacy, we provide a mathematic and two physics-based simulation examples, which demonstrate the adaptive surrogate modeling method can accurately predict the MTTF of the system in physics-based simulation model.