| Power equipment is the infrastructure of the whole power system,and in order to ensure good quality of power supply,it needs to run in normal condition for a long time.The substation plays the role of the top and bottom of the whole power system,and due to the advantages of non-contact detection of infrared detection technology,this technology is often used in substations for regular inspection of power equipment.The traditional method relies on the inspection personnel holding infrared equipment for inspection,due to the low efficiency of manual inspection,the number of inspectors is relatively small,in the work is also susceptible to the influence of the inspection personnel working time and subjective experience,resulting in not timely and effective judgment of the operating status of power equipment,can not meet the demand for real-time processing of infrared data of many substation equipment.In addition,the data for substation equipment infrared identification also has problems such as low image quality and relatively small number of data sets.Based on the above reasons,this paper develops the research on the infrared image recognition technology of substation equipment based on convolutional neural network,and the specific work is as follows:(1)Based on the causes of heat generation of power equipment and the principle of infrared imaging,common power equipment infrared detection methods are introduced and the advantages and disadvantages of these detection methods are described.The characteristics of YOLOv5 s target detection model are introduced,which lays the theoretical foundation for the target recognition of infrared images of four types of substation equipment below.(2)Four kinds of substation equipment infrared image datasets,such as insulators,bushings,10 k V switchgear copper rows,and 10 k V switchgear control circuit power switches,were obtained by reviewing various literatures,annotating these datasets,expanding the datasets by changing brightness,rotating angle,mirroring,etc.,and using data enhancement methods such as Mosaic,Simple Copy-paste,Mixup,and Cut Mix to finally generate usable four kinds of substation equipment infrared image datasets.(3)Optimize the infrared image recognition model of substation equipment based on YOLOv5 s by introducing Shuffle Net V2 in the Backbone of YOLOv5 s using channel mixing and washing for model lightweighting to further reduce its number of parameters and complexity while maintaining the excellent performance of the network model.Based on the lightweight model,CBAM,SENet,and ECA attention mechanism modules are introduced to the same location of Backbone to improve the detection performance of the model,respectively.(4)Experiments are conducted on the optimized model.The experimental results show that the model with the introduction of light weighting decreases significantly in terms of the number of parameters,54.6% compared to the YOLOv5 s model,and the computational speed of the model is improved and the m AP is decreased.The experimental analysis of the attention mechanism shows that the models with the introduction of CBAM and SENet attention mechanism have an improvement in m AP parameters relative to the lightweight model,but both are lower than the YOLOv5 s model,and the model with the introduction of ECA attention mechanism has the best effect,and the m AP is improved by 0.7% compared with YOLOv5 s.In conclusion,this paper optimizes the YOLOv5 s model by introducing channel blending and attention mechanism,and conducts experiments on the acquired substation equipment infrared image dataset.The experimental results show that the optimized YOLOv5s＿ECA model works optimally and achieves high detection accuracy while having faster detection speed,which verifies the effectiveness and feasibility of this paper’s convolutional neural network-based substation equipment infrared image recognition technology and lays the foundation for achieving intelligent detection and recognition of power equipment infrared images. |
PDF Full Text Request