Fault Diagnosis Of Smart Meters Based On Few-Shot Learning

Electricity anomalies have caused huge economic losses to grid companies for a long time.The abnormal diagnosis of power metering is of utmost importance to ensure the proper functioning of the meter,and a hot spot for the electric power staff.Since the announcement of the smart grid plan in China,smart meters have become popular in place of traditional electronic meters and the advanced metering reading system(AMR)has been continuously improved.The massive accumulation of data and real-time information have provided new opportunities for electricity metering abnormal diagnosis.Remote anomalies Diagnosing gradually replaces manual on-site diagnosis as an important direction for realizing grid automation operations.At present,the diagnosis of power metering anomalies is mainly based on rule-based feature engineering and expert systems.With the rapid development of electric meters and the rapid changes in the environment of power supply and consuming,the limited and lagging rules have become the bottleneck to improve the accuracy of abnormal diagnosis.More and more scholars have applied big data analysis technology to the electricity field.There are 2 challenges must be solved to build a data-driven diagnostic model: one is the lack of labeled samples,and the second is how to perceive unknown anomalies.In this paper,we apply a few-shot learning method based on metrics to the power field and design an end-to-end metrological abnormality diagnosis model.Firstly,we transform the original power data into a visual model via appropriate data processing and embed the knowledge of the field of abnormal diagnosis,which makes it easier for the multi-layer artificial neural network to extract the relevant feature from original data.Then we construct a classification network based on few-shot learning to identify known anomalies.Finally,this paper designs and implements an additional network for confirming the prediction results of the superior classification network.Through the cascade of the classification network and the additional network,a power metering anomaly diagnosis model obtains the ability to sense unknown categories.Experiments prove that the diagnostic model proposed in this paper has excellent performance in the abnormal diagnosis tasks of multiple test scenarios.Compared with the traditional remote diagnosis method,the abnormal diagnosis model doesn't need complex feature engineering and can identify rare abnormal types and unknown abnormal types with very few cases more precisely.In addition,the power data model conversion method proposed in this paper provide a reference for deep learning applications in the electricity field.