Data-driven Modeling in Digital Twin

Data-driven Modeling in Digital Twin

According to thelenComprehensiveReviewDigital2022

• Statistical models
  • Statistical models for dynamic system identification (P2V). Models: time-based (AR, ARMA, ARIMA, SSI), non-linear (NARX, NARMAX), frequency (FDD)
  • Statistical methods for degradation modeling in predictive maintenance (V2P). Models: single failure (Markov models, semi-Markov models, Wiener process, Poisson process, inverse Gaussian process, Gamma process, accelerated life testing (ALT) model), multiple failures (multiphase stochastic process, piece-wise Markov process, copula functions, system-level ALT), Bayesian model updating methods with Markov chain. (see: thelenComprehensiveReviewDigital2022)
• Machine Learning models
  • Conventional ML: feed-forward NN, SVM, random forests, Gaussian process regression
    • For system identification, combined with statistical methods (GP-NARX, NARX-net)
  • Deep learning model
    • For surrogate modeling. Algorithm: LSTM, CNN-LSTM
    • For system identification
      • DL is used to learn the input-output relationship directly. Algorithm: LSTM
      • DL is used to learn the state-space model of a system. Algorithm: autoencoder

Data-driven modeling: reasons

• the underlying physics is too complicated or is not fully understood
• the physics is understood and can be modeled using available software, but the simulation is too computationally expensive or time-consuming to be useful in a digital twin

Data-driven modeling: classes

• Data-driven models for degradation modeling
• Data-driven surrogate models. Surrogate modeling is typically constructed and optimized using computer simulation data for any type of physical simulation at any timescale.
• Data-driven models for dynamic system identification using sensor measurement. System identification mainly uses online sensor monitoring data and/or offline experimental data collected to tune the specific model’s parameter(s).

Data-driven modeling: issues

• Model selection
• Uncertainty quantity (UQ) of ML models
• Data collection