基于自监督学习的变电站缺陷检测

为了在使用少量标注样本情况下提升变电站设备缺陷检测精度,提出一种基于自监督模型SimSiam的改进缺陷检测算法.不同于原始SimSiam,改进后的算法无需使用标志性图像(iconic images),如ImageNet数据集,而是直接利用非标志性图像(non-iconic images)如COCO数据集进行对比学习,并在下游的缺陷检测任务上获得可媲美有监督方法的性能.通过在投影层(projection head)和预测层(prediction head)中使用全卷积网络和空间注意力模块来代替MLP,保留高维特征的空间结构及局部信息;同时在计算相似度前先对特征图进行均值池化以得到特征向量,并对特征向量进行归一化以计算欧氏距离,从而改进了自监督对比学习的损失函数.实验结果表明该算法能充分利用非标志性图像进行对比学习,并在只标注少量样本的条件下提升变电站设备缺陷检测的精度,在表计表盘破损、挂空悬浮物、鸟巢、呼吸器硅胶变色及箱门闭合异常等5类缺陷检测任务上mAP达到83.84%.

Substation Defect Detection Based on Self-supervised Learning

To improve the precision of equipment defect detection of substations on the premise of only a small number of labeled samples used, this study proposes an improved defect detection algorithm based on the self-supervised model SimSiam. Unlike the original SimSiam, the improved algorithm directly utilizes non-iconic images, such as those in the dataset COCO, rather than using iconic images, like those in the dataset ImageNet, for contrastive learning and achieves performance comparable to that of any supervised methods in downstream defect detection tasks. By replacing multi-layer perception (MLP) networks with fully convolutional networks and spatial attention modules in the projection and prediction heads, the proposed algorithm preserves the spatial structure and local information of high-dimensional features. Furthermore, the output feature map is mean-pooled before similarity is calculated to obtain the eigenvector, which is then normalized to calculate the Euclidean distance and further modify the loss function of self-supervised contrastive learning. The experimental results show that the improved algorithm can make full use of non-iconic images for contrastive learning and improve the precision of equipment defect detection of substations on the premise of labeling only a small number of samples. Its mean average precision (mAP) reaches 83.84% in the detection of five types of defects, namely, broken meter dials, hanging suspended substances, nests, respirator silicone discoloration, and abnormal closure of the box door.