The performance of a pure data-driven machine learning model highly depends on the quality and quantity of collected and the type of systems. There is no guarantee that the collected data can reveal the whole system. Meanwhile, the learning process by simply minimizing the deviations between predictions and data may violate the underlying physics. Therefore, an intrinsic shortcoming of DBM is that it cannot be easily parameterized, or extended to situations that the model has not been exposed to. Additionally, some practical issues like the noisy data and uncertain operation conditions can further jeopardize the validity of the data-based machine learning model.

In order to increase the applicability of the machine-learning model, our lab has been devoted to integrating physics to customize the machine-learning model for a specific problem. There are multiple methods to take advantage of the physics and dynamics of the system. By introducing the Lagrangian and modifying the construction of machine learning, the Lagrangian neural network is employed to model high dimension system. The PST method and SOD/POD method leverage the dynamics of the system to filter the raw data and extract more presentative features. Adaptive modeling uses physics to parameterize the system for tracking the varying parameters. Based on the analysis of the rotor model, we measure graphical data instead of one-dimensional data to capture the behavior of the rotor and extract more powerful data in rotor diagnostics. This method is called the imagine-capturing method. More details and applications can be found below.

•   Hybrid Modeling of a Multidimensional Coupled Nonlinear System With Integration of Hamiltonian Mechanics

•   Physics-informed Machine Learning for Modeling Multidimensional Dynamics

•   Adaptive Modeling

•  POD-SOD methods for signal processing

•   Periventricular Leukomalacia (PVL) Prediction

•   Arrhythmia Detection

•   Gear Diagnostics

•  Injected Diagnostic Signal

•  Image processing for diagnostics