一种语义分割引导的呼吸器油位计异常检测方法

A SEMANTIC SEGMENTATION GUIDED ABNORMALITY DETECTION METHOD FOR RESPIRATOR OIL LEVEL GAUGE

针对变电站异常数据量少、难以支撑深度神经网络训练的问题,以变压器的呼吸器油位计识别问题为例,提出一种融合虚拟数据和语义分割的呼吸器油位计异常检测方法,目标是提升变压器状态监测的识别能力。具体地,构建融合语义分割和分类器的异常检测模型,通过油位计图像的语义分割获取具有空间语义信息的特征,同时采用分类器获得准确的异常状态类别;提出融合虚拟数据和真实数据的模型训练方法,通过调整两种类型数据的比例,实现在少量真实数据下模型的有效训练。实验结果表明,虚拟数据能够有效缓解真实数据不足的问题,且该方法相比采用DCNN模型的直接分类,能够达到更高的异常检测精度。

To tackle the scarcity of abnormal data in the substation for training deep neural networks,we particularly consider the identification of transformer respirator oil level gauge in this paper,and propose a semantic segmentation guided abnormality detection method with synthetic data.Specifically,an anomaly detection model integrating semantic segmentation and classifier was constructed.The semantic segmentation model could extract rich features with spatial semantic information of the input image,and meanwhile the classifier could predict accurately abnormal state category.Moreover,we proposed a training method integrating synthetic data and real data.Through adjusting the proportion of the two types of data,we could effectively train the model even with a small amount of real data.The experimental results show that the synthetic data can effectively alleviate the insufficient problem of real data,and the proposed method can achieve higher accuracy than directly using DCNN classification model.