| The surge in the number of substations make it difficult for manual inspection to cover the needs of maintenance,so it is of great significance to carry out intelligent inspection.At present,the main means of intelligent detection of substation defects is to collect data through visible light,infrared and ultraviolet imagers,then make analysis and judgment on this basis.The surface defects,including breakage in dial plate and the missing in umbrellas of silicon rubber insulator,will lead to security concerns.Timely discovery and rectification of them can effectively improve the operation stability of the substation.However,the complex-imaging environment of substation makes a low identification accuracy.To address it,this thesis carries out an in-depth study on the image recognition methods involved in the above question.The specific work is as follows:Firstly,the image acquisition process of substation equipment is easily influenced by environmental lighting factors,which leads to image contour blurring and information loss.To address this question,a low illumination image enhancement algorithm based on improved hue mapping is proposed.In order to improve the image quality and reduce the number of false missed detection,a low illumination image enhancement algorithm based on improved hue mapping was proposed.The algorithm firstly decomposes the image into high and low frequency sub-bands by dual-tree complex wavelet transform,then the decomposed high and low frequency sub-bands are processed by guided filtering algorithm and improved hue mapping method,and finally,the image with higher resolution is synthesized by inverse dual-tree complex wavelet transform.On this basis,identification data set of substation equipment is constructed through image annotation,image enhancement and other operations.Second,in view of the problems of low recognition rate caused by the similar appearance and structure of substation equipment,complex background and the inclusion of multiple objects with significant scale differences in the same image,an improved SSD identification method for substation equipment was proposed.To ensure extraction of multi-scale features and enhance the effectiveness of large-scale features,this method adopts deep residual network to replace VGG16 in traditional SSD as the backbone network.At the same time,the strategy of default box generation is optimized,and K-means clustering method is adopted to obtain the target default box with significant size difference in accordance with this thesis.On this basis,transfer learning and data enhancement techniques are applied to train the network model of substation equipment identification to improve the robustness of the network and make up for the impact of the small number of samples on network training.Compared with the traditional SSD algorithm,the proposed algorithm improves the accuracy by 1.43%,the recall rate by 42.24%,and the harmonic mean by36.26%.Third,to solve the problem of unbalanced image samples in surface defect recognition of substation equipment,a surface defect recognition algorithm based on deep convolution generated adversative network is proposed.The method takes the dial as an example to study the surface defect identification.Firstly,the generated model is trained by using the dial image under normal state,then the image to be detected is input to the generated model,and the input image is compared with the generated image to realize the judgment of the dial surface state,and realize the surface defect recognition under the imbalance of samples.The detection accuracy rate can reach 94.6%.Finally,the software system of surface defect identification for substation equipment is designed and developed,and the application test is carried out.It is verified that the design of the surface defect identification system is reasonable and feasible,and a good man-machine interaction is realized.It can be applied easily and quickly in laboratory testing and substation practical application. |
PDF Full Text Request