基于原型智能网络的电网检修状态异常检测模型

A Maintenance State Anomaly Detection Model for Power Grids Based on Prototype Intelligent Networks

电网装置通常安装在户外,会受到大量污染。污染物积聚在电网装置中,可能会引起短路并导致停电。为了提高电网的可靠性,利用计算机视觉技术实现自动化电网检修状态异常的检测。提出一种基于原型(Prototype)智能网络的电网检修状态异常检测模型(Proto-PGNet),为自动化电网检修状态异常检测提供辅助决策。由于现有电网检修数据集包含的不同背景图像数量有限,如何使模型更具泛化性是一个挑战。Proto-PGNet模型在最后一个密集层上不进行凸优化,以保持逆向推理过程对图像分类的作用。逆向推理过程可以排除输入图像中的错误类别,可以用少量且具有不同背景的图像进行分类。Proto-PGNet模型与其他先进模型进行对比实验,结果表明Proto-PGNet模型明显优于其他模型。其中,以VGG-19为网络骨架时,Proto-PGNet的准确率达到了97.22%,比最先进的Ps-PGNet模型的准确率提高了4.17%。

Grid installations are usually installed outdoors and are subject to a large amount of pollution.When pollutants build up in grid installations,they can cause short circuits and lead to power outages.In order to improve the reliability of power grids,computer vision technology is used to automatically detect the abnormal state of power grid maintenance.A maintenance state anomaly detection model is proposed for power grids based on prototype intelligent networks(Proto-PGNet),which provides auxiliary decision-making for automatic power grid maintenance state anomaly detection.Due to the limited number of different background images contained in the existing grid maintenance data sets,how to make the model more generalized is a challenge.The Proto-PGNet model does not perform convex optimization on the last dense layer to maintain the effect of backward inference on image classification.The backward inference process can remove the wrong categories in the input images,so it can be classified with a small number of images with different backgrounds.The comparison experiments between the Proto-PGNet model and other advanced models show that Proto-PGNet is obviously superior to other models.When VGG-19 is used as the network backbone,the accuracy of Proto-PGNet reaches 97.22%,which is 4.17%higher than that of the most advanced Ps-PGNet model.