Optimized YOLOv7-tiny model for smoke detection in power transmission lines

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Fire incidents near power transmission lines present significant safety risks.
• Existing smoke detection models lack accuracy and speed due to complex scenarios.
• Fast and accurate smoke detection is vital for power system operational safety.

Purpose of the Study:

• To develop an improved smoke detection model for high-voltage power transmission lines.
• To enhance detection accuracy and speed compared to existing methods.
• To ensure the safety and reliability of power system operations.

Main Methods:

• Constructed a specialized dataset using particle system-generated smoke composited into real scenes.
• Introduced parameter-free attention modules and Spd-Conv (Space-to-depth Convolution) into YOLOv7-tiny.
• Employed transfer learning by pre-training on synthesized data and fine-tuning on real-world scenarios.

Main Results:

• Achieved a 2.61% increase in mean Average Precision (mAP) for smoke detection.
• Improved precision by 2.26% and recall by 7.25% compared to the original YOLOv7-tiny model.
• Demonstrated superior accuracy and speed against other object detection models and improved performance on the Figlib dataset.

Conclusions:

• The improved YOLOv7-tiny model offers enhanced smoke detection capabilities for power transmission lines.
• The methodology effectively addresses limitations in existing smoke detection systems.
• This advancement contributes to improved safety and operational integrity of power infrastructure.