Bio-inspired multi-dimensional deep fusion learning for predicting dynamical aerospace propulsion systems.

Area of Science:

• Aerospace Engineering
• Artificial Intelligence
• Dynamical Systems Analysis

Background:

• Forecasting dynamical systems is crucial for aerospace safety.
• Existing methods often focus on global trends, neglecting local periodicity in time series data.
• Aerospace propulsion data exhibits distinct dynamical periodicity over limited timeframes.

Purpose of the Study:

• To develop a deep learning model, TimeWaves, capable of capturing both global trends and local variations in dynamical systems.
• To enhance the prediction accuracy for time series data with inherent periodicity.
• To address the challenge of predicting rocket combustion instability.

Main Methods:

• Developed TimeWaves, a deep learning model utilizing 3D spectrum-oriented interval extraction.
• Integrated Fourier and Wavelet analyses via a shared parameter fusion algorithm.
• Employed a dual-way learning workflow with TwinBlock for efficient multi-scale feature perception.

Main Results:

• TimeWaves effectively captures both global trends and local variations in time series data.
• The model demonstrates robust performance in predicting rocket combustion instability.
• Achieved enhanced perception of dynamical multi-scale features at reduced computational cost.

Conclusions:

• TimeWaves offers a novel and effective approach for forecasting dynamical systems with periodicity.
• The model significantly improves the prediction of critical aerospace phenomena like rocket combustion instability.
• This deep learning framework provides a reliable tool for enhancing aerospace mission safety and efficiency.