英文摘要 | The safe and reliable operation of high-voltage transmission lines is directly related
to the development of national economy and the stability of people’s lives, so transmission lines must be inspected regularly. However, the common manual inspection has
high safety risk and low inspection efciency. At present, robot autonomous inspections are the main development direction of smart grids, and autonomous environment
perception and defect detection are the key to autonomous inspections. However, there
are many difculties in autonomous environment perception and defect detection. It
is difcult to achieve automatic identifcation of the variable scale of the wire when
the robot approaches the wire; the positioning of multi-posture insulators under complex background is not accurate; the detection accuracy of small-scale defects of power
equipment is low. Therefore, studying the autonomous perception of lines and insulators, and high-precision defect detection of power equipment have important theoretical
signifcance and engineering value. This dissertation focuses on power line pose detection, insulator three-dimensional reconstruction, insulator segmentation, and typical
equipment defect detection in a power transmission environment. The main work of the
dissertation is as follows:
Firstly, in order to solve the problem of wire scale change recognition, an efcient
parallel branch network for real-time overhead power line segmentation is proposed,
which realizes the high accuracy positioning of variable scale power lines. Semantic
branch and spatial branch are used to obtain semantic features and spatial features respectively. The asymmetric factorized depth-wise bottleneck (AFDB) module with low
parameters efciently completes short-range feature extraction. On this basis, a classifer module with skip connections is designed, and the performance of power line
segmentation is improved through the fusion of short-range features and long-range
features. In addition, a line pose detection method based on connected components and
moments is used to detect the principal axis direction and centroid of the power line.
Secondly, aiming at the difculty in obtaining actual samples of insulators, an insulator reconstruction method with scaling volume-view supervision from single and
multiple images is proposed, which realizes the expansion of insulator samples. Upsampling layer is used to increase the error of the original reconstruction result, thereby
improving the 3D reconstruction performance. On this basis, a three-dimensional reconstructionmethodofinsulatorsformultipleimagesbasedontheattentionmechanism
is further proposed. The recurrent neural network is used to realize the perspective
selection in the three-dimensional reconstruction process, and the accuracy of the reconstruction results is further enhanced. Based on this method, the two-dimensional
contour image of the insulator can be obtained by projection for the insulator sample
augmentation, which solves the problem of data acquisition difculty.
Thirdly, for the complex background of insulator images and the different scales
of insulators, an insulator segmentation method based on the conditional generative adversarial network is proposed to achieve high-accuracy segmentation of insulators in
complex backgrounds. The generator and discriminator are composed of a convolution
group based on decomposition and convolution. The generator effectively extracts rich
low-order features, and the discriminator obtains high-order features of the image. In
order to further reduce the network parameters, expand the receptive feld and improve
the segmentation accuracy, the residual depth-wise separable convolution block is proposed,andthemoduleisintegratedintotheU-Netframeworktobuildagenerator,which
improves the multi-scale feature fusion capability and then enhance the segmentation
performance.
Fourthly,aimingattheproblemofthesmallsizeofpowerequipmentdefect,asmall
targetdefectdetectionmethodbasedonmulti-scalefeaturefusionisproposedtoachieve
efcient and accurate detection of small-scale defects. By adding a batch normalization
layer to the convolutional block attention module, the distribution of the network input
is guaranteed to be the same, and the influence of different channels on the feature map
is strengthened. The multi-scale feature fusion module improves the ability to extract
semantic features of small-scale targets by cascading multi-scale feature maps at one
time. Finally, the multi-branch detection module is used to suppress the low-quality
bounding box and improve the accuracy of defect detection.Finally, the research work of this dissertation is summarized, and future research
plans are proposed.
|
修改评论