A Self-Adaptive 1D Convolutional Neural Network for Flight-State Identification.

Area of Science:

• Aerospace Engineering
• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Wing vibrations result from coupled aerodynamic-mechanical responses during flight.
• Identifying flight states (angle of attack, airspeed) from complex vibration signals is challenging.
• Self-sensing wings offer potential for real-time structural health monitoring and flight state determination.

Purpose of the Study:

• To develop a novel method for accurate flight-state identification using structural vibration data from a self-sensing wing.
• To automatically extract relevant features from complex vibration signals.
• To enhance the self-awareness capabilities of intelligent air vehicles.

Main Methods:

• A one-dimension convolutional neural network (CNN) was developed for feature extraction.
• Dual-tree complex-wavelet packet transformation decomposed vibration signals into frequency bands.
• A grey-wolf optimizer optimized key CNN parameters.
• Reconstructed sub-signals were combined for multichannel CNN input.

Main Results:

• The proposed CNN method demonstrated high accuracy in identifying flight states.
• The method showed robustness compared to standard deep-learning approaches.
• Feature extraction was automated, reducing manual effort.

Conclusions:

• The developed CNN method provides an effective approach for flight-state identification from wing vibrations.
• This research offers new insights for the development of self-aware intelligent air vehicles.
• The combination of signal decomposition, CNN, and grey-wolf optimization shows significant promise.