| Distribution cabinets in a large number of wiring attachments and management of electrical equipment has a significant impact on the good operation,due to the number of wiring types of a wide range,there may be misconnections,omissions and other phenomena,which can seriously threaten the safe operation of substations,and even lead to equipment burn,bringing economic losses to enterprises,and for substations secondary circuit screen cabinet wiring construction process mainly relies on manual operation in accordance with the drawings,not only efficiency low,there is also the possibility of false detection,leakage.To address these problems,this paper proposes a Mask R-CNN-based automatic identification method for connector parameters,building a neural network model for optimised target detection,which can accurately identify the terminal ranking,terminal number and cable number and match them,assisting the operation and maintenance personnel to detect the connection of cables,reducing the error rate of human eye detection and improving detection efficiency.The main research content and results are as follows:1)Collection and labelling of cable connector parameter data sets.The data set is the most basic part of the intelligent detection of the secondary circuit of the power grid,and since relatively few data sets can be collected at present,this paper uses mobile phone photography to collect the distribution cabinet site,collates more than two thousand pictures of cable joint parameters,and after a lot of manual operations,the target area in each image is boxed and text marked,and several proofreading and adjustments are carried out to establish a cable joint The data set of cable connector parameters was established.2)The Mask RCNN-based target region localization network was improved.The Deformable ROI pooling network was integrated into the Mask RCNN,and an adaptive offset was introduced to improve the ROI Align layer,resulting in a cable numbered region localization network that can accurately detect small targets in images.After comparing with the mainstream detection models,it is demonstrated that the improved method in this paper has the highest localisation rate of 91.85% in the cable numbering region.3)An optimised method of fusing the difference operator of Gaussian function with vertical projection is proposed to cut the text region into individual characters,which effectively reduces the noise impact of the data set.By combining edge expansion techniques with cubic spline interpolation,the individual characters are normalised to create a cable parametric character dataset that is clearer and less distorted.4)Based on the characteristics of the cable connector parameter dataset,a cable parameter character recognition network called Classify-CNN is proposed and compared with existing detection methods.At the same time,a visual interface was designed to compare with the cable wiring standard library while detecting,which can extract the cables with wrong connections,laying the foundation for the unmanned detection of distribution cabinets. |
