Research On Intelligent Detection Of Protective Wear For Power Grid Operation Based On Deep Learning

In recent years,in view of the traditional manual-based standardized safety dressing inspection means,the detection results are easily affected by the supervisor’s state and other factors.And it is difficult to achieve real-time supervision of the whole process of production operations.The intelligent monitoring technology of protective equipment wearing compliance based on the combination of machine vision and artificial intelligence technology has gradually become a research hotspot in this field.Based on the above background,this thesis taked the compliance detection of work clothes and insulated gloves for operators in the process of power grid operation as a research example.Based on the actual video images collected in the field and the research results in related fields,this thesis conducted an in-depth research on the intelligent detection technology of wearing compliance for labor protection gear based on deep learning.The main research work and research results are as follows:(1)According to the actual environmental characteristics of the power grid operation site,a overall process design scheme based deep learning for the compliance testing of power grid protective equipment wearing was proposed.The whole detection process was divided into three main steps in sequence: operator identification,image extraction of detected area from operators,and image detection of key area.At the same time,according to the needs of the research work in this thesis,the production of relevant data sets was completed by using the video images collected at the job site.(2)In the research of automatic identification of specific target personnel,due to the interference of factors such as complex actual work site environment and small difference in characteristics between operators and supervisors,the direct application of YOLOv5 s model for operator identification has high false detection rate and missed detection rate.By adopting improvement strategies such as introducing an attention module in the backbone network and optimizing the loss function,an improved operator identification method based on YOLOv5 s was proposed.Experimental results show that the improved model can further improve the accuracy of target person recognition and reduce the missed detection rate while meeting the real-time requirements.(3)In the research of image extraction from the operator’s area to be detected,firstly,the body parts that should be covered by work clothes and insulated gloves through empirical analysis,and they are set as the regions of interest.On this basis,an improved image multi-category semantic segmentation algorithm based on UNet++ was proposed.The experimental results show that introducing data augmentation operations before detection,using residual network(Res Net34)to improve the standard Encoder unit of the original UNet++ model,and combining BCE and lovasz-Softmax to improve the loss function can further improve the segmentation accuracy of the target region.It lays a foundation for the subsequent compliance detection for the region of interest.In research process,the proposed model algorithms were implemented in Python language,and the experimental scheme was designed and verified according to the specific research content.According to the results of all the experimental data analysis,the proposed deep learning-based protective wear compliance detection method for power grid operations can better meet the application requirements in real application situations and has certain promotion value.