油纸绝缘沿面放电是导致油浸式变压器绝缘劣化的重要原因，放电碳痕的精准分割对于放电程度判断及绝缘劣化趋势分析具有重要价值。针对树枝状放电碳痕边缘复杂、尺寸差异大、油渍浸润干扰等严峻挑战，提出了实现碳痕精准分割的MCH-UNet模型（multi-head coordinate-attention hardswish UNet）。针对油渍浸润干扰、补光不足等导致的像素值集中、对比度偏弱等问题，采用自动色彩均衡算法（automatic color enhancement, ACE）进行碳痕样本的图像增强；设计多头注意力卷积模块（multi-head attention convolution, MHAC），代替原有的常规卷积模块，在降低模型复杂度的同时增强不同尺度碳痕特征的融合效果；为缓解编码器与解码器之间的语义差距，提高碳痕特征激活强度和分割完整度，采用CA注意力机制替代原有的跳跃连接；采用Hardswish激活函数替代原有的ReLU函数，改善模型训练时的梯度传播，提高模型对复杂边界的解析能力。试验结果表明，与UNet、UNet ++、UneXt、MALUNet、Att-UNet、EGE-UNet等模型相比，MCH-UNet具有最佳的分割效果；相比于基准模型UNet，MCH-UNet将平均交并比MIoU、像素准确率PA和精确率PE分别提升了1.83%、1.05%、4.33%；同时，MCH-UNet的模型参数量和GFLOPs分别为0.121 × 10⁶和0.074，具有较好的轻量化优势。研究结果可为油浸式变压器内放电碳痕的精准分割和油纸绝缘健康状况评估提供技术支撑。

Surface discharge of oil-paper insulation is an important cause of insulation degradation in the oil-immersed transformer, and the accurate segmentation of discharge carbon traces is of great value for the judgment of discharge degree and the analysis of insulation degradation trend. In order to meet the challenges of complex edges, large size differences and oil infiltration interference of dendritic discharge carbon traces, MCH-UNet (multi-head coordinate-attention hardswish UNet) was proposed to realize the accurate segmentation of carbon traces. To address the problems of concentrated pixel values and weak contrast caused by the interference of oil stain infiltration and insufficient supplementary light, ACE (automatic color enhancement) algorithm was adopted for image enhancement of carbon trace samples. was proposed MHAC (multi-head attention convolution) module was designed to replace the conventional convolution module, for enhancing the fusion effect of carbon trace features at different scales while reducing the model complexity. In order to alleviate the semantic gap between the encoder and the decoder, and to improve the activation strength of the carbon trace features and the segmentation completeness, the CA attention mechanism was used instead of the original jump connection. The hardswish activation function was used instead of the original ReLU function to improve the gradient propagation during model training, and improve the ability to resolve complex boundaries. The experimental results showed that MCH-UNet had the best segmentation effect compared with UNet, UNet ++, UneXt, MALUNet, Att-UNet, EGE-UNet and other models. Compared with the benchmark model UNet, MCH-UNet improved MIoU (mean intersection over union), PA(pixel accuracy), and PE(precision rate) by 1.83%, 1.05%, and 4.33%, respectively. At the same time, the model parametric quantities and GFLOPs of MCH-UNet are 0.121 × 10⁶ and 0.074, respectively, which has a better lightweight advantage. The results could provide technical support for the accurate segmentation of discharge carbon trace and the assessment of oil-paper insulation health condition in the oil-immersed transformer.