一种基于注意力机制的绝缘子状态反馈认知方法

针对已有绝缘子状态识别模型,以及细节识别深层网络开环认知模式和损失函数泛化能力不足的缺陷,模仿人工巡检模式,即实时评估认知结果可信度自寻优调节多尺度图像知识空间决策,提出一种基于注意力机制的绝缘子状态反馈认知方法。首先,针对预处理后的绝缘子图像,设计自适应尺度堆叠的卷积神经网络架构,使得网络输入由整体图像缩放至细节局部区域,每个尺度的网络共享相同的架构具有不同的参数,确保不同分辨率输入的可区分能力并为下一尺度生成一个细节注意区域。其次,随机配置网络面向多个尺度特征,建立具有强泛化能力的绝缘子状态分类准则。再次,构建类间分类和类内尺度间排序损失函数优化注意力网络,较前次预测生成更高置信度得分排名。最后,借鉴闭环控制思想,定义广义误差熵性能指标实时评测绝缘子不确定状态认知结果可信度,动态调节网络尺度等级,实现不确定认知结果约束下的特征空间自优化调节和分类准则重构,反馈再认知绝缘子状态。实验结果表明了本文方法与其他网络架构相比,增强了模型的泛化能力,提升了模型的认知精度。

A feedback cognition method of insulator state based on attention mechanism

To overcome the drawbacks of the existing insulator state recognition models, open-loop cognitive mode and insufficient generalization ability of loss function for detailed recognition deep network, in this paper a feedback cognition method of insulator state is proposed based on attention mechanism in imitation of human inspection mode, i.e. real-time evaluation of reliability of cognitive results and self-optimizing and regulation of the multi-scale image knowledge space. Firstly, for the pre-processed insulator image, a stacked convolutional neural network with adaptive scale architecture is designed, which enables the network input to be scaled from the overall image to the detailed area. Each scaled network shares the same architecture with different parameters to ensure the discriminative ability of different resolution inputs and generate a detailed attention area for the next scale. Secondly, for multiple scale features, stochastic configuration network (SCN) builds the classification criterion of the insulator states with universal approximation ability. Thirdly, an inter-class classification loss function and an intra-class ranking loss function are constructed to optimize the attention network, which generates a higher confidence score ranking than the previous prediction. Finally, learning from closed-loop control idea, the generalized error entropy performance index is defined to evaluate the reliability of the uncertain cognition results of insulator states in real time.The network scale level is dynamically regulated to realize the self-optimizing regulation of the feature space and the reconstruction of the classification criteria based on the constraint of the uncertain detection results, which enables insulator states to be re-recognized with feedback mechanism. Experimental results show that compared with other network architectures, the proposed method enhances the generalization ability and improves the cognition accuracy of the model.