Research On Visual Fault Classification Of Electric Meter Based On Deep Learning

As an important end device for smart grid marketing business,electricity consumption information and energy distribution,smart meters play an important role in supporting the stable operation of smart grid.Visual faults such as display failure,surface breakage,burnout and missing seals can occur during the operation of smart meters.With the explosive growth of the number of smart meters,the traditional manual way of troubleshooting and processing information has been difficult to meet the demand of efficient operation and maintenance.The layout structure of meters of different fault types is highly similar,the visual differences of fault features are small,and the discriminative features are difficult to be mined,while there are significant differences in the number of samples of various types of faults,which easily lead to the decision results of fault classification models biased towards the majority class and bring great challenges for accurate identification of visual faults of meters.Machine vision technology based on deep learning is an effective means to mine image information,automatically identify visual faults of electricity meters,and improve the intelligence level of operation and maintenance operations.Therefore,in this paper,we conduct a technical study on the classification task of meter visual faults based on deep learning,and the main work is as follows:First,a meter visual fault classification method oriented to the small inter-class variance feature is studied.The mainstream fine-grained image classification methods mine discriminative features through multigranularity jigsaw and masking mechanisms,but the unreasonable jigsaw division easily leads to incomplete or redundant features within the jigsaw patches,while the masking mechanism of constant occlusion background discards the effective information in non-discriminative features.To address the above problems,a fine-grained image classification method based on discriminative granularity adaptive setting and fading mask is proposed:First,an attention map is constructed to present the importance distribution of target features,and according to the outline of important feature regions in the map,the granularity values corresponding to the division of the jigsaw mechanism are calculated,and the discriminative granularity values reflecting the typical size of target features are clustered and mined,according to which the granularity of the division of the jigsaw is adaptively set,effectively preserving the integrity of semantic features within the jigsaw patches while reducing redundant information.On this basis,the non-discriminative feature regions are mined according to the target feature importance distribution,and masks are applied to it.The mask probability is attenuated in 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 classification.Finally,the multi-grain size features are fused for classification.Experimental comparisons with typical fine-grained image classification methods on public and meter datasets demonstrate the effectiveness and advancement of the proposed method.Secondly,a method for generating visual fault images of electricity meters under the characteristics of imbalance in the number of multi-class fault samples is studied.The current mainstream generative adversarial networks do not sufficiently learn the minority class image features,and the discriminator is prone to overfitting,making it difficult to generate an effective gradient to optimize the generator.To address the above problems,a conditional generative adversarial network based on branchenhanced discrimination is proposed:First,category embedding encoding is introduced to increase the distance of latent feature vectors of each class of samples in the latent layer space,which facilitates the generative adversarial network to learn different distribution patterns among highly similar classes of samples.Second,a global augmented discriminator and a local augmented discriminator are designed to discriminate from global and local perspectives,respectively,to improve the diversity of discriminating perspectives and input samples,and to prevent the discriminator from being trained perfectly too early and thus overfitting to a few classes of samples,resulting in too small or disappearing gradients,and the generator cannot be trained effectively.On this basis,the original samples and the generated samples are co-augmented to prevent the generator from learning the distribution of the augmented samples while improving the generalization ability of the discriminator.Experimental comparisons with typical image generation methods on public and metered datasets demonstrate the effectiveness and advancement of the proposed method.Finally,a framework for meter information extraction and visual fault detection in multi-meter boxes is studied.The current mainstream object detection models have limited ability to mine the micro features of the target.Direct detection of meter faults by target detection models is prone to false detection.To address the above problems,a framework for meter information extraction and visual fault detection in multi-meter boxes is proposed:Firstly,a target detection model is used for meter type identification and screenshots according to the regression coordinates,and the intercepted single meter image is input into the meter visual fault classification model and barcode detection model to obtain the fault prediction category and identity information of the meter image.Secondly,a meter location identification algorithm is designed to identify the row and column position of each meter according to the coordinates output from the object detection model.Finally,The fault information of the meter is correlated with the meter location information,identity information and meter location information and fed back to the operation and maintenance personnel to improve efficiency and maintain the stable operation of power business.