Transfer Learning Based Method for Frequency Response Model Updating with Insufficient Data.

Area of Science:

• Mechanical Engineering
• Computational Mechanics

Background:

• Finite element model updating requires precise vibration feature extraction.
• Collecting sufficient vibration data for frequency response (FR) model updating is time-consuming.
• Accurate feature extraction with limited data presents a significant challenge.

Purpose of the Study:

• To develop a novel transfer learning approach for finite element model updating with small datasets.
• To leverage existing fault diagnosis datasets for enhanced model updating.
• To reduce the dependency on extensive data collection in FR model updating.

Main Methods:

• A transfer learning network with source and target domain feature extractors was proposed.
• A domain adaptation method was employed to align features in a shared space.
• A readily available fault diagnosis dataset was used as ancillary knowledge.

Main Results:

• The transfer learning method significantly reduced sample amount dependency.
• The updated model demonstrated superior accuracy compared to methods without transfer learning on small datasets.
• Validation through dynamic responses outside the training set confirmed the model's reliability.

Conclusions:

• Transfer learning offers a viable solution for accurate finite element model updating with limited vibration data.
• The proposed domain adaptation technique effectively bridges cross-domain feature discrepancies.
• This approach enhances the efficiency and accuracy of FR model updating processes.