SAVSDN: A Scene-Aware Video Spark Detection Network for Aero Engine Intelligent Test

Area of Science:

• Aerospace Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Sparks in aero engine chambers, caused by carbon deposits, lean flames, or damaged parts, are currently detected manually.
• Existing object detectors struggle with high-precision spark detection due to similar features between sparks and interference.

Purpose of the Study:

• To develop an automated, high-precision spark detection system for aero engines.
• To address the limitations of current object detection methods in distinguishing sparks from interference.
• To create a valuable dataset for training and evaluating spark detection models.

Main Methods:

• Proposed a scene-aware spark detection network (SAVSDN) using an information fusion-based cascading video codec-image object detector structure.
• Emphasized extracting spatio-temporal features from adjacent frames to minimize over-detection.
• Developed a Gaussian function-based method for generating simulated aero engine spark images and introduced the SAES dataset.

Main Results:

• SAVSDN demonstrated the ability to learn distinct spatio-temporal features differentiating sparks from interference.
• The proposed simulation method successfully generated realistic spark images.
• SAVSDN significantly outperformed state-of-the-art detection models across five key metrics in experimental evaluations.

Conclusions:

• The developed SAVSDN offers a robust and accurate solution for automated spark detection in aero engines.
• The introduction of the SAES dataset provides a crucial resource for advancing research in this domain.
• SAVSDN's scene-aware approach effectively overcomes the challenges posed by interference in spark detection.