基于Mask R-CNN与改进BP神经网络联合算法的变压器套管红外热故障诊断

为解决传统图像类算法在变压器套管状态诊断时存在的效率低、准确度不高以及复杂背景下变电设备目标识别困难等问题,提出了将Mask R-CNN与改进BP神经网络相结合的套管红外图像状态诊断方法。首先,利用Mask R-CNN解决套管红外图像背景复杂时分割困难的问题;其次,基于灰度特征的特征量提取方案,实现对红外伪彩图特征量的提取;最后,引入粒子群优化BP神经网络（PSO-BP）算法对变压器套管特征进行分类识别。实验结果表明,该方法对红外图像中套管的运行状态具有较好的检测效果,对套管中介质损耗故障、接头故障和漏油故障的故障诊断准确率分别可达100.0%、88.9%和96.3%,平均准确率达到93.518%,优于传统BP算法和支撑向量机（SVM）算法。

Thermal Fault Diagnosis of the Bushing Infrared Images Based on Mask R-CNN and Improved BP Neural Network Joint Algorithm

To solve the problems of low efficiency, low accuracy and difficult target recognition of substation equipment under complex backgrounds in the traditional image algorithm in the state diagnosis of transformer bushings, this paper proposes a combination of diagnosis method of tube infrared image state with Mask R-CNN and improved BP neural network.First, Mask R-CNN is used to solve the problem of difficult segmentation when the background of the casing infrared image is complex;secondly, the feature extraction scheme based on grey features realizes the extraction of infrared pseudo-color image features;finally, the particle swarm optimization BP (PSO-BP) in neural network algorithm is introduced to classify and identify the transformer bushing features.The results show that this method has a good detection effect on the running state of the casing in the infrared image, and the fault diagnosis accuracy of the dielectric loss fault, joint fault and oil leakage fault in the casing can reach 100%, 88.9% and 96.3% respectively, the average accuracy reaches 93.518%, which is superior to the traditional BP algorithm and SVM algorithm.