An Unsupervised Situation Awareness Framework for UAV Sensor Data Fusion Enabled by a Stabilized Deep Variational...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Effective situation awareness in autonomous systems requires processing high-dimensional sensor data.
• Deep generative models face challenges with training instability and multi-modal data in UAVs.
• Existing methods struggle with complex, non-linear sensor time-series data.

Purpose of the Study:

• To propose a novel unsupervised sensor data processing framework for UAVs.
• To address training instability and multi-modal distribution representation challenges.
• To enable robust feature extraction from complex UAV sensor data.

Main Methods:

• Developed a deep generative model, VAE-WRBM-MDN, for non-linear time-series sensor data.
• Utilized Weighted-uncertainty Restricted Boltzmann Machines (WRBM) for stable layer-wise pre-training.
• Integrated a Mixture Density Network (MDN) for accurate multi-modal distribution reconstruction.

Main Results:

• Achieved 95.69% classification accuracy in identifying situational patterns.
• Demonstrated stable learning and convergence, overcoming limitations of standard Variational Autoencoders (VAEs).
• Successfully reconstructed complex, multi-modal conditional distributions of sensor readings.

Conclusions:

• The VAE-WRBM-MDN framework offers a robust solution for UAV sensor data processing.
• This approach enhances real-time intelligent sensing and raw data interpretation.
• The proposed method provides enabling technology for advanced autonomous systems.