Hybrid methods for prognosis of mechanical components have the potential of improving remaining useful life (RUL) estimations. Hybrid methods combine physics-based and data-driven methods to diagnose faults and predict when failures will occur. In this work, we propose a hybrid framework that estimates the RUL from routine maintenance inspection data and condition monitoring data.
This hybrid framework is applied to turbomachine rotor blades. Blade tip timing (BTT) measurements are used for condition monitoring. The least squares spectral analysis (LSSA) method is used to find the natural frequency. The natural frequency is a function of the blade's health state and is used to infer the crack length in the blade’s root. To accommodate for artificial rotational stiffening, an interpolation model of the blade's Campbell diagram, generated from a finite element model, is used.
In the proposed methodology, we use an ensemble physics-based model that serves as a prior probability density function for a Gaussian process regression (GPR) model. The predictive distribution of the GPR model is constructed by conditioning the physics-based model on the observed data from routine maintenance inspections. We also compare this hybrid diagnosis model to the physics-based technique and other data-driven methods. The hybrid diagnosis model outperformed the physics-based method and data-driven methods.
The unscented Kalman filter (UKF) is used to estimate and forecast the evolution of the crack length over time. Paris’ law coefficients are used as hidden latent variables in a hybrid degradation model. In the diagnosis and prognosis phases, we use the unscented transform to efficiently approximate the probability density functions of the crack length and the crack growth law parameters.
Finally, we compare the applied hybrid model to a purely physics-based model and show that the hybrid model outperforms the physics-based model in predicting the length of a crack. We demonstrate that the hybrid model increases the accuracy and precision of RUL prediction from physics-based models with 60% on average.