Research On Power Equipment Defect Detection Based On Multiple Feature Fusion And Deep Steady-State Learning

The cost and difficulty of traditional manual inspection are high,and rapid positioning of power equipment is a prerequisite for researching power equipment defect detection.Therefore,this article will further study and apply power vision technology based on image data of power equipment.The main research work is as follows:1.An EOSHHO-SVM power equipment defect detection algorithm based on multiple feature fusion is proposed.In order to obtain objective and complete features of power equipment based on the classification and description requirements of normal and fault states of power equipment,multiple feature extraction methods such as SIFT algorithm are used and feature fusion is performed.The maximum information coefficient(MIC)is combined to remove the correlation of the fusion features,and the main features obtained are used for classification training of support vector machines(SVM),At the same time,the main parameters of SVM classifier are optimized and improved through the improved Elite Opposition Learning Strategy based Harris Hawks Optimization(EOSHHO)algorithm.The experimental results show that the proposed algorithm’s defect classification test scores are2.5% and 7.5% higher than HHO-SVM and PSO-SVM algorithms,respectively;2.A causal feature fusion algorithm for power equipment defect detection based on steady-state learning is proposed.In view of the difficulties in optimizing the current YOLOv5 network model and the large amount of computation,the paper first adopts the EOSHHO optimization algorithm to replace the Adam optimizer,and the deep separable convolution to replace the ordinary convolution;Secondly,using an improved feature pyramid structure,the attention mechanism is introduced;Thirdly,replace the SPPF module with a pyramid pooled PPM module;Finally,in order to remove irrelevant features and false associations,the paper uses a method that combines steady state learning and target detection.The experimental results show that compared with YOLOv5 s algorithm,the proposed algorithm improves the overall average accuracy and recall rate by 1.8% and 3.9%,respectively,while ensuring that the model weight size and reasoning time are basically unchanged and increased.