Transformer abnormal heat accurate identification method based on AHIPDNet.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Transformer operation is critical for power system stability.
• Accurate detection of abnormal transformer heat is essential for preventing failures.
• Existing methods may struggle with small targets or precise localization of heat anomalies.

Purpose of the Study:

• To enhance the accuracy and reliability of abnormal transformer heat recognition.
• To develop an improved target detection algorithm for transformer health monitoring.
• To improve the identification of small abnormal heat spots in transformers.

Main Methods:

• An improved YOLOv8 target detection algorithm was developed.
• SPD-Conv convolutional layers were integrated to enhance small target recognition.
• HAT hybrid attention and Dyhead multi-attention mechanisms were incorporated for focused hot spot identification.
• The proposed method is named AHIPDNet (Abnormal Heat Identification and Positioning Detection Network).

Main Results:

• The AHIPDNet model achieved a recognition accuracy of 93.1%.
• The integrated attention mechanisms effectively focused on abnormal hot spot information.
• The SPD-Conv layer improved the recognition of small abnormal heat targets.
• The model demonstrated accurate recognition and localization capabilities in field conditions.

Conclusions:

• The improved YOLOv8 model, AHIPDNet, significantly enhances abnormal transformer heat recognition accuracy.
• The combination of advanced convolutional layers and attention mechanisms provides a robust solution for transformer monitoring.
• This method contributes to improved power system safety and stability through effective fault detection.