Research And Implementation Of Image Anomaly Detection For Power Grid Equipment Based On Deep Learning

With the development of the power grid industry,the scale of the power grid system has become more and more large.In order to prevent economic losses,the problem of equipment safety detection in substations has become particularly important.However,human-based detection brings a series of drawbacks and excessive resource consumption.Therefore,it is a significant subject to apply artificial intelligence on abnormal detection for the images of power grid equipments.In actual scenarios,the difference among abnormal classes of power grid equipment is small and difficult to discern.Besides,the algorithm needs to be integrated into the system,which requires high computing speed.To solve these problems,we propose a deep learning based anomaly detection method for power grid equipment image:an efficient and multi semantic device anomaly detection algorithm.This algorithm contains multiple high-level semantic extraction modules.Meanwhile,we compress the anomaly detection algorithm,which can be deployed end to end and its prediction speed reaches real time.First,we design the basic feature enhancement network,which combines the multi-scale feature semantics in each sub-module and enhances the model’s semantic extraction ability and generalization ability.Second,we design a new anomaly detection loss function,which defines a new object candidate box calculation method,improving the ability of the model for positioning objects.Third,we design a new anomaly fine-grained classification module,which incorporates a layered bilinear pooling method and boosts the accuracy for recognizing.Fourth,we use the method of model compression to shrink the model,which greatly reduces the amount of model calculation under the premise of small loss of detection accuracy.Finally,in order to make the algorithm applicable to the actual scenarios,we also constructed an abnormal image data set to support model training and testing.In addition,in order to make the coverage of power grid equipment anomaly detection more comprehensive,we use image registration auxiliary algorithm to perform auxiliary detection on undefined power grid equipment anomaly categories.At the same time,we design the grid equipment anomaly detection scheduling algorithm,which is responsible for fusing the prediction results of the detection algorithm and the image registration auxiliary algorithm.After using the scheduling algorithm,the accuracy of the anomaly detection algorithm can reach 95%.Finally,we establish a prototype system for grid equipment anomaly detection based on web applications,which integrates grid equipment anomaly detection algorithms.It can detect anomalies in real-time monitoring images and provide early warnings.In order to ensure the high reliability of equipment abnormality detection,the staff conducts a review according to the early warnings.