Dynamic Vision-Based Non-Contact Rotating Machine Fault Diagnosis with EViT.

Area of Science:

• Robotics and Automation
• Computer Vision
• Machine Learning

Background:

• Event-based cameras provide high temporal resolution and dynamic range for capturing mechanical phenomena.
• Conventional imaging systems struggle with complex dynamic vision data from these sensors.
• Machine vibration sensing and fault diagnosis require advanced non-contact methods.

Purpose of the Study:

• To develop a non-contact machine fault diagnosis method using dynamic vision data.
• To introduce an effective computational framework for processing asynchronous event streams.
• To enhance predictive maintenance in smart manufacturing environments.

Main Methods:

• Proposed the Eagle Vision Transformer (EViT) architecture with Bi-Fovea Self-Attention and Bi-Fovea Feedforward Networks.
• Developed a computational framework to process asynchronous event streams while preserving temporal precision.
• Utilized multi-scale spatiotemporal feature analysis and adaptive learning.

Main Results:

• Demonstrated exceptional fault diagnosis performance across diverse operational scenarios.
• Successfully processed complex dynamic vision data from event-based cameras.
• Validated the method using practical condition monitoring data from rotating machines.

Conclusions:

• The proposed EViT method offers a new paradigm for non-contact machinery fault diagnosis.
• This approach effectively bridges bio-inspired vision processing with industrial monitoring needs.
• The findings provide valuable insights for predictive maintenance applications.