【作者】 曹勇；

【导师】 徐造林； 张柏礼；

【作者基本信息】 东南大学 ， 计算机科学与技术， 2022， 硕士

【摘要】 金属锈蚀会对变电设备的正常运转造成威胁,需要及时发现并处理,以避免形成安全隐患。采用机器人替代人工巡检的变电设备锈蚀缺陷检测能够降低管理成本,并避免人为疏忽,更好地实现对变电设备的实时监控。实际中,机器人按照线路巡检,将采集到的设备图像传送至锈蚀缺陷检测模块进行检测,因此,该方案非常依赖部署在边缘设备即机器人上的锈蚀缺陷检测能力。相较于其他的锈蚀检测技术,基于卷积神经网络的检测技术不依赖于人工定义的特征而成为了目前的研究热点。然而,现有针对变电设备锈蚀缺陷检测的研究都仅将原有目标检测模型迁移过来,或者针对特征提取网络进行一定的改进,而忽略了锈蚀缺陷样本与普通图像样本之间的差异。对此,本文首先针对金属锈蚀缺陷的特点提出一种分层嵌套标注方法;然后,针对特征图全局信息表达能力不强的问题,将自注意力机制融入特征提取网络;最后,考虑到边缘设备有限的计算资源,采用PQF(Permute,Quantize and Fine-tune)网络压缩方法进行检测模型压缩,使其更加易于部署在边缘设备上。综上,本文提出一种基于卷积神经网络的变电设备锈蚀缺陷检测方案,具体研究内容包括:(1)金属锈蚀缺陷形状的不规则性和可拆分性导致在训练样本标注时,标注者面临着很多歧义,难以实现标注过程的标准化和标注结果的一致性,本文提出一种新的训练样本分层嵌套标注方法。这种标注方法一方面不存在标注歧义,易于统一和标准化,容易形成稳定的标注质量;另一方面突出了锈蚀特征,并一定程度上增加了标注为锈蚀的样本数量,实现了数据增强。(2)虽然卷积神经网络能够通过叠加层数的方式扩大感受野,但相比于注意力机制,全局信息的融合能力仍有不足。为此,本文在特征提取网络后,加入自注意力机制提高特征图的全局信息表达能力。在大规模图像数据集Image Net上的实验证明该方法能够提高特征图所包含的全局信息,进而提高模型的分类能力。进一步实验表明:融合自注意力机制后,可进一步提升变电设备锈蚀缺陷检测模型的性能。(3)鉴于原始检测模型的体积较大,压缩后再将其部署于边缘设备是一种优化方向。PQF网络压缩方法经过排列组合、矢量量化和模型微调三个步骤压缩网络模型,实验证明该压缩方法能够在变电设备锈蚀缺陷检测模型性能略微下降的情况下,大幅降低检测模型对计算资源的需求。因此,采用该压缩方法后,使得较为复杂庞大的金属锈蚀检测模型在性能、空间都受限的边缘设备上的部署更为便利。更多还原

【Abstract】 Metal corrosion is harmful to the power equipment,which is necessary to find and handle corrosion in time to avoid potential safety hazards.Corrosion detection of power equipment using robots instead of engineers can reduce the management cost and avert human negligence,and realize the real-time monitoring of power equipment.The robots collect equipment information according to the inspection path and send it to corrosion detection system for detection.Therefore,this method depends heavily on the corrosion detection ability deploying on edge equipment.Compared with other corrosion detection methods,the method based on convolutional neural network without artificially defined features,and become a research hotspot at present.However,the existing research on corrosion detection of power equipment only migrates original object detection models or improves the feature extraction networks,but ignores the difference between corrosion samples and common image samples.In this thesis,hierarchical annotation method is proposed for corrosion features;then,for the weak global information expression ability of feature map,self-attention mechanism is integrated into feature extraction network;finally,considering the limited computing resources of edge devices,PQF(Permute,Quantize and Fine-tune)method is adopted to compress the detection model to easier deploy it on edge devices.To sum up,this thesis proposes a corrosion detection system of power equipment based on convolutional neural network.The specific research contents include:(1)The irregularity and detachability of metal corrosion makes annotators confront ambiguity and uncertainty in the labeling process.This thesis proposes a novel hierarchical annotation method.This method can easily produce unified annotation results without ambiguity.And it highlights the corrosion features and increases the number of ground truth,realizing data augmentation.(2)Convolutional neural network expands its receptive field by stacking layers,but compared with attention mechanism,the global information expression ability is still insufficient.Therefore,behind the feature extraction network,self-attention mechanism is added to improve the global information expression ability of feature map.Experiments on large-scale image dataset Image Net show that this method improves the global information contained in feature map,then improves the classification ability of model.Further experiments show that merging self-attention further improves the performance of corrosion detection model.(3)Considering the large volume of original detection model,it is an optimizing direction to deploy model on edge devices after compression.The PQF method compresses network model by three steps: permute,quantize and fine-tune.Experiments show that this method greatly reduces the demand for computing resources of detection model when the performance decreases slightly.Therefore,after adopting PQF,more complex corrosion detection model becomes convenient to deploy it on edge equipment with limited performance and space.更多还原