Loading...

Physics-Informed Neural Network in Fault Diagnosis of Rotating Machines

Ahmadi Ghouchan Atigh, Nastaran | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: English
  3. Document No: 57797 (58)
  4. University: Sharif University of Technology
  5. Department: Science and Engineering
  6. Advisor(s): Behzad, Mehdi; Rohani Bastami, Abbas
  7. Abstract:
  8. Intelligent fault diagnosis (IFD) involves the application of machine learning (ML) theories to machine fault diagnosis. Vibration analysis is a common technique used for diagnostics and prognostics. A hybrid method that integrates both physics-based and data-driven approaches could offer an effective strategy for monitoring health. In this research, the artificial neural network (ANN) and convolutional neural network (CNN) algorithms, which are common models, have been used. Then, based on the results obtained, developing a physics-informed neural networks (PINN) model, more specifically physics-informed convolutional neural network (PICNN), is proposed. The algorithm used for PICNN includes a main CNN (called CNN A) that is a vibration-data-driven model, and a complementary CNN (called CNN B) that specifically learns from GMF, its sidebands, and its second and third harmonics. The PICNN model improves the detection accuracy through a customized loss function in which the model updates the weights when the predictions of CNN A differ (based on the healthy and damaged classes). Technically, in the customized loss function, the weights in CNN A are modified by considering the weights of CNN B. The proposed physics-informed deep learning approach is validated using 1) Data from the gearbox on a laboratory test stand at Sharif University of Technology, and 2) a Planetary gearbox prognostics dataset from the University of Pretoria. The validation results are obtained using PICNN, CNN, and ANN models. Based on the results achieved with the two different tests, it is concluded that the proposed PICNN model has higher accuracy, and more stability compared to the CNN and ANN models
  9. Keywords:
  10. Intelligent Fault Detection ; Machine Learning ; Artificial Neural Network ; Convolutional Neural Network ; Physics Informed Neural Network ; Physics-Informed Convolutional Neural Network (PICNN)

 Digital Object List

 Bookmark

No TOC