基于鉴别性粒度自适应设定和衰退掩码的智能电表可视故障分类方法

Classification of Smart Meter Visual Faults Based on Discriminative Granularity Adaptive Setting and Fading Mask

实现智能电表可视故障的精准检测,对电网计量现场的高效运维至关重要。不同故障电表的布局结构高度相似,其特征呈现类间方差小的特点,细粒度图像分类方法是此条件下挖掘鉴别性特征的有效手段。在目前主流研究中,混淆拼图机制引导模型学习固定粒度的特征,但容易导致拼图块内的特征冗余或不完整。掩码机制通过恒定遮挡非鉴别性特征区域来突出鉴别性特征,但忽视了该区域中有助于分类的信息。该文提出了一种基于鉴别性粒度自适应设定和衰退掩码的智能电表可视故障分类方法。首先,对训练图像构建注意力图,以呈现目标特征的重要性分布,将图中重要特征的轮廓尺寸转换为等效粒度值并进行聚类挖掘,获得反映目标特征尺寸特点的鉴别性粒度值,据此自适应设定拼图的划分粒度,有效保留拼图块内语义特征完整性的同时减少冗余信息;在此基础上,根据目标特征重要性分布挖掘非鉴别性特征区域,并对其施加掩码,在迭代训练中衰减掩码概率,逐步降低对该区域的遮挡程度,引导模型学习此区域中有助于分类的特征信息;最后结合渐进式多粒度特征引导学习框架,融合不同粒度的特征信息以完成分类。在多个权威公开的细粒度图像分类数据集和智能电表可视故障数据集开展大量实验,与10种典型细粒度图像分类方法对比,验证了所提方法在准确率等指标上的先进性。

Achieving accurate detection of the visual faults in the smart meters is crucial for the efficient operation and maintenance at the grid metering sites.Since different faulty meters have the highly similar layout structures with the features of small inter-class variance,the fine-grained image classification methods are effective means to mine the discriminative features under this condition.In the current mainstream researches,the confusion jigsaw mechanism is used to guide the model to learn the fixed granularity features,which tends to result in the redundant or incomplete features within the jigsaw patches.The masking mechanism is adopted to highlight the discriminative features by its constant masking of the non-discriminative feature regions,which ignores the information in the region that helps the classification.In this paper,a visual fault classification for smart meters based on the discriminative granularity adaptive setting and the fading mask is proposed.First,an attention map is constructed to present the importance distribution of the target features.The outlines of the important features in the map are transformed into the granularity values corresponding to the division of the jigsaw mechanism,and the discriminative granularity values reflecting the typical size of the target features are clustered and mined.According to this granularity of the division of the jigsaw,the integrity of the semantic features are effectively preserved within the jigsaw patches while reducing the redundant information.On this basis,the non-discriminative feature regions are mined according to the target feature importance distribution,and masks are applied to them.The mask probability is attenuated in the iterative training to gradually reduce the occlusion degree of this region and guide the model to learn the feature information in this region that helps the classification.Finally,aprogressivemulti-granularity feature-guided learning framework is combined to fuse the feature information of different granularities to complete the classification.Extensive experiments are conducted on several authoritative and publicly available fine-grained image classification datasets and the smart meter visual fault dataset,and the proposed method is compared with 10 typical fine-grained image classification methods to verify its advancedness in terms of accuracy and the other metrics.