基于红外图像的变电设备热缺陷自调整残差网络诊断模型

Self-adjusting Residual Network Diagnosis Model for Substation Equipment Thermal Defects Based on Infrared Image

针对部分设备外形相似、热缺陷状态区分度不高、模型参数过多等导致基于红外图像的变电设备缺陷状态诊断模型复杂、准确度不高等问题,构建了变电设备红外图像缺陷状态多分类数据集,提出了一种基于红外图像的热缺陷自调整残差网络诊断方法。首先,通过卷积核分解技术优化残差网络基础结构,减少模型参数数量;然后采用多尺度卷积特征融合方法,对网络浅层和深层产生的判定特征进行融合,提高对缺陷状态的识别准确率;最后,提出基于约束改进的贝叶斯优化算法,在准确率和网络体积两约束条件下,实现卷积核个数、网络深度等超参数的自调整,获取性能最优的轻量化诊断模型。研究结果表明:所提改进模型的状态识别准确率达到94.53%,比Alexnet、残差网络(residual network,Resnet)等经典网络高出约3%,可为电力设备的故障诊断提供参考。

A self-adjusting residual network diagnosis model for substation equipment defects based on infrared images,is proposed.In this model,substation equipment defect state diagrams are used as the data source to solve the problems of inaccurate diagnosis caused by the similar appearance in some devices,low discrimination in various defect states,and excessive parameters in some models.Firstly,the residual network infrastructure is optimized by the convolution kernel decomposition technology to reduce the number of model parameters.Secondly,the multi-scale convolution feature fusion method is adopted to fuse the features of low layers and deep layers so as to improve the recognition accuracy of the defect state diagrams.Finally,a Bayesian optimization algorithm based on constraint improvement is proposed.Under the constraints of verification set and network volume,the self-adjustment of hyper parameters,such as the number of convolution kernels and the depth of the network,is realized to obtain a lightweight diagnostic model with optimal performance.The results show that the state recognition accuracy of improved model reaches 94.53%,which is about 3%higher than that of Alexnet and Resnet.It can provide a reference for fault diagnosis of power equipment.