A Deep Learning-Based Unbalanced Force Identification of the Hypergravity Centrifuge.

Area of Science:

• Engineering
• Physics
• Machine Learning

Background:

• Unbalanced forces in hypergravity centrifuges impact operational safety and experimental accuracy.
• Accurate identification is crucial for mitigating these effects and improving model testing.

Purpose of the Study:

• To develop a deep learning-based model for accurate and quantitative identification of unbalanced forces.
• To enhance the model's performance by integrating handcrafted features with a Residual Network (ResNet).
• To optimize the model for imbalanced datasets and validate its effectiveness.

Main Methods:

• A feature fusion framework combining ResNet with handcrafted features was established.
• Loss function optimization was performed for imbalanced datasets.
• A shaft oscillation dataset was created using the ZJU-400 hypergravity centrifuge with artificially added unbalanced mass.
• The deep learning model was trained and evaluated on this dataset.

Main Results:

• The proposed model demonstrated superior accuracy and stability compared to benchmark models.
• Mean Absolute Error (MAE) was reduced by 15%–51%, and Root Mean Square Error (RMSE) by 22%–55% on the test dataset.
• Continuous identification during speed-up showed high accuracy and stability, outperforming traditional methods significantly.

Conclusions:

• The deep learning-based unbalanced force identification model offers a robust solution for hypergravity centrifuge applications.
• The method provides valuable guidance for counterweight adjustments, enhancing unit stability.
• This approach significantly improves the safety and reliability of hypergravity operations and model testing.