A Small Object Detection Method for Oil Leakage Defects in Substations Based on Improved Faster-RCNN

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Substation equipment safety is crucial for reliable power transmission.
• Current oil leakage detection methods struggle with small defects and lack intelligent robotic integration.
• Intelligent inspection robots are essential for efficient substation maintenance.

Purpose of the Study:

• To develop an accurate small object detection method for oil leakage in substations.
• To integrate this method with intelligent inspection robots for enhanced substation monitoring.
• To provide actionable maintenance recommendations for oil leakage incidents.

Main Methods:

• Modified Faster R-CNN model using Resnet-101 feature extraction.
• Implemented modifications include canceling downsampling and replacing large convolutional kernels with smaller ones to preserve information, especially for small objects.
• Integrated the detection model with an intelligent inspection robot and developed a decision-making scheme for maintenance recommendations.

Main Results:

• The proposed FRRNet101-c model demonstrated superior performance in oil leakage detection compared to baseline models.
• Achieved a 6.3% improvement in Mean Average Precision (mAP) for overall detection.
• Showcased a significant 12% improvement specifically in detecting small oil leakage objects.

Conclusions:

• The FRRNet101-c model offers a highly effective solution for detecting oil leakage defects in substation equipment, particularly small ones.
• The integration with intelligent inspection robots enhances substation inspection capabilities and supports timely maintenance decisions.
• This approach contributes to ensuring the longevity of equipment and the stable operation of power systems.