改进SSD模型的绝缘子自爆故障检测

insulators self-explosion fault detection based on improved SSD model

绝缘子是输电线路中重要的电力设备,直接影响电力系统的稳定运行,而传统的目标检测技术难以准确、高效地检测绝缘子自爆故障。为解决该问题的同时提高自爆故障检测的精度和召回率,文中提出一种基于轻量级目标预测层的多尺度特征融合SSD(PL-MFSSD)模型。首先,使用深度可分离卷积代替传统卷积操作生成目标预测特征层,提升模型的检测效率;其次,在Conv_fc7和Conv8_2特征层中进行多尺度特征融合,将目标的浅层结构信息和深层语义信息充分融合,并在目标预测特征层末端增加残差网络,缓解训练过程中的梯度消失问题;最后,通过K-means聚类算法优化锚框的宽高比,使其更符合绝缘子和自爆故障的尺寸。在NVIDIA GTX1080实验环境下,PL-MFSSD模型的mAP指标为0.941,自爆故障的召回率达到0.967,推理速度为49.62 f/s。实验结果表明,与其他检测模型相比,PL-MFSSD模型对自爆故障的召回率有大幅度提升,可有效检测绝缘子自爆故障。

The insulator is an important power equipment in transmission line,which directly affects the stable operation of power system,but the traditional object detection technology is difficult to detect insulator self-explosion fault accurately and efficiently.A multi-scale feature fusion SSD model based on lightweight target prediction feature layer(PL-MFSSD)is proposed to resolve above problem and improve the detection accuracy and recall of self-explosion fault.The depth separable convolution is used to replace the traditional convolution operation to generate object prediction feature layer and improve the detection efficiency of the model.The multi-scale feature fusion is carried out in Conv_fc7 and Conv8_2 feature layers,which fully fuses the shallow structural information and deep semantic information of the object.The residual network is added at the end of the object prediction feature layer to alleviate the problem of gradient disappearance in the process of training.The aspect ratio of anchor box is optimized by means of the K-means clustering algorithm to make it more in line with the size of insulator and selfexplosion fault.In NVIDIA GTX1080 experimental environment,the mAP of PL-MFSSD model is 0.941,the recall of self-explosion fault is 0.967,and the inference speed is 49.62 f/s.the experimental results show that in comparison with other detection models,PL-MFSSD model is able to greatly improve the recall of self-explosion fault,and can effectively detect insulator self-explosion fault.