Research On Object Recognition Algorithm For Corrosion Area Detection Of Power Equipment

The safety monitoring of power equipment is the guarantee for the stable operation of power systems,and the detection of rust area is a very important part in the fault detection of power equipment,which can effectively prevent the occurrence of electrical fires.The safety monitoring of power systems has always relied on manual inspections,which are inefficient and low in security.The increasingly developed image recognition technology can judge whether there are rust and other faults in the power equipment based on the image information,but the rusted areas of the power equipment generally have complex modes.The practical problems of many aspects such as difficulty in feature extraction lead to a low recognition rate in actual detection.How to speed up the detection speed under the condition of improving the recognition rate has also become the main research direction of scholars at home and abroad.At present,the accuracy rate of rusted image classification algorithms has been low,the memory space required by the target detection algorithm is large,and the required computing power is strong,which cannot be real-time monitoring.To solve these problems,this paper combined the target detection algorithm with multi-scale features,depthwise separable convolution and attention model to improve the rust target location and recognition algorithm,and achieved good detection results.The main research contents and results are summarized as follows:(1)Research on image classification algorithm based on multi-scale feature fusion method.Aiming at the problems of poor classification of rust image and low recognition rate by convolutional neural network,this paper proposes a rust image classification algorithm based on multi-scale feature fusion method.Using the excellent feature extraction ability of deep convolutional neural network,combined with the high-level semantic information of deep-scale and the fine-grained information of shallow-scale to enhance the feature.Then the feature reduction and fusion are performed by PCA,which can remove the invalid feature information.Finally,SVM is used instead of the fully connected layer for classification.The experimental results show that the proposed algorithm can obtain higher detection accuracy and prove the effectiveness and superiority of the proposed algorithm.(2)Research on target detection algorithm based on lightweight neural network.Aiming at the problems of large target size and slow calculation of current target detection algorithms,this paper proposes a rust target detection algorithm based on lightweight neural network.The use of depthwise separable convolution instead of standard convolution guarantees a certain detection accuracy while the parameter amount is greatly reduced.Secondly,a splicing combination and an upsampling strategy of adding channel attention are used to compensate for the accuracy due to the reduction of the parameter amount.According to the experiment,the algorithm can maintain good detection results with less parameter and faster detection speed.(3)Research on target detection algorithm based on attention model.In view of the above-mentioned accuracy degradation caused by the lightweight model,this paper proposes a rusted area detection algorithm for the cascaded dual attention model.Combining the spatial attention model and the channel attention model,a cascading dual attention model is constructed by splicing,and the model can be added as a sub-module in any target detection algorithm.The experimental results show that the proposed algorithm can greatly improve the accuracy of detection while maintaining high-speed detection.It also proves the practicability and effectiveness of the proposed algorithm in the electrical system.