ML Algorithms

Physics-Informed AI

  • A physics-informed approach was developed (defect-free model was built)

  • Physics-Informed features were extracted using residual and cross-sample-entropy analysis

  • Random forest model integrated with SHAP algorithm was implemented to rank the extracted features


Sensor Fusion Using CNN-LSTM Network

  • What is the best way to integrate data from various sources?

  • A integrative deep learning approach to fuse data from different sensors

  • The algorithms is based on two parts:

        • CNN are used to extract features from different time serieses

        • LSTM are used to rank and select the extracted features from different sources

  • Publications: IEEE

Embedding Dimension Diagnostic Method

  • How can we measure a system's dimensionality?

  • A novel method was developed during my doctoral work

  • Based on the information in the embedding dimension of a given signal

  • Capable of distinguishing various nonlinear responses within the system (chaos, multi-periodic, etc.)

  • Applied to gear-train setup to identify cracks

Transfer Learning for Nonlinear Pendulum

  • Develop a transfer learning approach to update fault identifier when the system changes

  • Structural change in the system

          • A large amount of data available before the structural change

          • Limited data after the structural change

  • Change in operation scenario

          • A large amount of data for previous operation scenario

          • Limited data for new operation scenario

Phase Space Topology Family of Methods

  • Industrial machinery applications reported nonlinear phenomena such multi-periodic, quasi periodic and chaotic that were originating from defects or due to their nonlinear nature.

  • How can we characterize system response and get insight using nonlinear dynamics?

  • A novel family of methods was developed during the my doctoral work

  • Based on extracting information from the phase space domain

  • Generalized on various range on real mechanical and electrical systems

  • A patent was filled

  • Publications: VETOMAC| JVET | Grant | Patent

Ranked Recurrence Diagnostic Method

  • Based on integrating mutual information and recurrence quantification analysis

  • Recurrence plots reveals all the times when a dynamic system visits roughly the same area in the phase space (repeating different kinds of behavior)

Publications: JSEA