基于多尺度网络的绝缘子自曝状态智能认知方法研究

Research on intelligent cognition method of insulator self-blast state based on multi-scale network

针对已有绝缘子状态识别模型,以及深层网络尺度和交叉熵损失函数的缺陷,仿照运维人员检修模式,即依据评测结果的可信度动态决策,基于多尺度网络构建了一种绝缘子自曝状态智能认知方法。首先,面向定位归一化化预处理后的绝缘子图像,基于ResNet-18增加不同结构的网络分支提高网络适应不同分辨率的能力,同时在网络末端添加多尺度信息融合模块;其次,随机配置网络面向多个尺度特征,构建了泛化的自曝状态分类认知准则;最后,为了评测自曝状态分类认知结果的可信度,基于定义的误差指标自调节多尺度网络架构,重构不确定认知结果约束下的特征向量和分类认知准则,以进行自曝状态再认知。实验结果显示,与其他方法相比,所提出的智能认知方法增强了模型的泛化能力和认知精度。

In view of the drawbacks of the existing insulator state recognition models,and the scale and softmax loss function of deep network,imitating the mode of personnel operation and maintenance,that is,dynamic decision-making based on the credibility of the evaluation results,this paper constructs an intelligent cognition method of insulator self-blast states based on the multi-scale network.Firstly,for the pre-processed insulator images with localization and normalization,based on ResNet-18,branches with different network structure are added to improve the network ability to adapt to different resolutions.At the same time,the multi-scale information fusion module is added at the end of the network.Secondly,facing multiple scale features,stochastic configuration network(SCN)constructs a generalized cognition criterion of self-blast state classification.Finally,in order to evaluate the credibility of the self-blast state cognition result,based on the defined error index,the multi-scale network architecture is self-adjusted to reconstruct the feature vector and classification cognition criterion under the constraint of the uncertain cognition result,which carries out the self-blast state renewal cognition.The experimental results show that the proposed intelligent cognition method enhances the generalization ability and cognition accuracy compared with other methods.