PSH‐YOLO: A Detection Method for Small‐Target Thermal Defects in Porcelain Insulators

ABSTRACT In recent years, infrared image‐based insulator defect detection technology has been widely applied in the field of online monitoring for power equipment due to its noncontact and high‐efficiency characteristics; however, existing algorithms still face issues such as insufficient detection accuracy and low computational efficiency in multistate insulator classification tasks, making it difficult to meet practical engineering requirements; to address these challenges, this paper proposes an improved small‐target multidefect detection algorithm Porcelain insulator Small‐target Heating defect detection You Only Look Once (PSH‐YOLO): based on YOLOv8, it employs a hybrid model of self‐attention and convolution to aggregate both convolutional and self‐attention features; then applies the MobileViT network to enhance the model's training speed and parameter efficiency, ensuring the overall lightweight nature of the model; additionally incorporates a bidirectional feature pyramid network to improve accuracy through multilevel feature pyramids and bidirectional information flow; finally, utilizes the Inner‐WIoU loss function to effectively reduce oscillations during training while further enhancing the model's accuracy; to obtain test data, this paper conducted infrared imaging experiments on defective insulators to capture images under varying conditions; experimental validation confirms that the proposed multidefect small‐target YOLO algorithm, PSH‐YOLO, achieves an average accuracy improvement of 6.17%, with Giga Floating‐point Operations Per Second reduced to 7.1, fulfilling the requirements for identifying small‐target insulator defects, while ablation and comparative studies demonstrate the effectiveness and superiority of the proposed algorithm.