Unsupervised Fault Detection on Unmanned Aerial Vehicles: Encoding and Thresholding Approach

Area of Science:

• Aerospace Engineering
• Artificial Intelligence
• Control Systems

Background:

• Unmanned Aerial Vehicles (UAVs) offer societal benefits but face safety challenges due to control component faults.
• Existing fault detection methods, like rule-based and supervised learning, have limitations including manual rule updates, extensive labeled data needs, and inability to detect novel faults.

Purpose of the Study:

• To develop an unsupervised fault detection model for UAVs that overcomes the limitations of prior approaches.
• To enhance UAV safety by accurately identifying known and unknown fault types.

Main Methods:

• Feature extraction from raw UAV flight logs.
• Development of a fault detection model integrating a stacked autoencoder and a classifier.
• Training the autoencoder on safe UAV flight data and using reconstruction loss to differentiate between safe and faulty states.

Main Results:

• The proposed stacked autoencoder model demonstrated effective fault detection capabilities.
• The approach successfully identified different types of UAV faults across two distinct datasets.
• The unsupervised learning method proved capable of detecting faults beyond those explicitly trained.

Conclusions:

• The unsupervised stacked autoencoder approach offers a robust solution for UAV fault detection, addressing limitations of previous methods.
• This model enhances UAV safety by reliably detecting both trained and novel fault types.
• The feature extraction and model design contribute to improved fault recognition in critical aerospace applications.