Detection and Quantitative Evaluation of Internal Bubble Defects in Basin Insulators Using Infrared Thermography

Internal bubble defects in basin insulators may cause power system failures, making accurate detection crucial. To address background noise interference, boundary blurring, and challenges in quantitative depth evaluation during detection, this study proposes a defect detection and quantitative evaluation method based on infrared thermography and deep learning. A detection system based on pulsed infrared thermal wave technology is constructed, which suppresses background noise and enhances defect features by optimizing thermal excitation parameters and image acquisition conditions. An improved Attention U-Net model, the Squeeze-and-Excitation Attention U-Net (SEA-UNet), is proposed by incorporating the Squeeze-and-Excitation Network (SENet) attention mechanism and a 2-layer 3×3 convolution module, which improves the ability to capture edge details, reduces boundary blurring, and enhances segmentation performance. Bubble diameter is quantified using a diameter calibration method, while defect depth is estimated via the Peak Second-order Derivative Time (PSDT) method. The proposed method can detect bubbles with a minimum diameter of 0.7 mm and identify bubbles at a depth of 2.0 mm. The model achieves a Mean Intersection over Union (MIoU) of 92.43%, Recall of 96.11%, and Precision of 95.71%, with average relative errors of 7.92% for diameter calibration and 9.50% for depth estimation. This study provides reliable technical support for the safe operation of power equipment.