CN-119671929-B - Power transmission line icing detection method based on wide-area perception dynamic convolution model

CN119671929BCN 119671929 BCN119671929 BCN 119671929BCN-119671929-B

Abstract

The invention discloses a power transmission line icing detection method based on a wide-area perception dynamic convolution model, which comprises the steps of firstly optimizing feature sharing among convolution kernels by using a Multi-dimensional attention module (Multi-Dimensional Attention Module, MDAM) to obtain detailed information of a small target more accurately, secondly providing a Thin-head feature extraction network fused with a light-level model (LIGHTWEIGHT MODEL, LM) to reserve hidden connection for gradually transmitting spatial feature information to a channel, and reducing underfitting generated by imbalance of positive and negative samples by using a loss function (Shape-NWD) combining a focused form cross ratio and normalization. Finally, experiments fully demonstrate the effectiveness of the method of the invention.

Inventors

ZHANG LU
KONG ZHIZHAN
HU PANFENG
CHEN SONGBO
ZHU MINGXI
PU LU
HAN YANHUA
LI WEI
ZHANG PENG
WANG WENSEN
WANG NAN
LIU XIAOBO

Assignees

国网陕西省电力有限公司电力科学研究院
国网(西安)环保技术中心有限公司

Dates

Publication Date: 20260508
Application Date: 20240929

Claims (6)

1. The method for detecting the icing of the power transmission line based on the wide-area perception dynamic convolution model is characterized by comprising the following steps of: Step 1, acquiring ice-covered images of the power transmission line in different environments through unmanned aerial vehicle inspection, expanding the obtained ice-covered images of the power transmission line, and taking the expanded ice-covered images of the power transmission line as a real insulator data set; step 2, dividing the ice coating data set of the power transmission line obtained in the step 1 into a training set and a verification set, wherein the ratio of the training set to the verification set is 9:1; step 3, labeling the training set in the step 2 for the ice coating type of the power transmission line by utilizing Labellmg software, labeling the training set by using a VOC format in image labeling, and storing the image information of the ice coating of the power transmission line after labeling to obtain an ice coating database of the power transmission line; Step 4, replacing the common convolution in the backbone network with an MDAM multidimensional attention module, wherein the MDAM extracts characteristic information from three different scales to obtain each characteristic map, and fusing the characteristic maps by using convolution and up-sampling operation; The method comprises the steps of merging a Thin-head feature extraction network of a lightweight model, introducing a mixed convolution RS Conv, including shuffling operation and DW Conv, calculating a weighted sum of neighbor node features of nodes by the RS Conv by using a learnable weight matrix, splicing the weighted sum with own node features, and then processing the spliced features by a nonlinear activation function to obtain updated node representation; the idea of Shape-IoU is fused into the normalized NWD to obtain Shape-NWD; The MDAM, the Thin-head feature extraction network fused with the light model and the Shape-NWD loss function are combined to form a MDAL-DETR target detection network; Step 5, predicting the verification set in the step 2 by utilizing the optimal weight network obtained after training in the step 4 to obtain a corresponding power transmission line icing detection result; And 6, evaluating the detection result obtained in the step 5 by adopting an accuracy rate, a recall rate and a mAP value to obtain the efficiency of detecting the icing of the power transmission line by each model and the practicability of the improved algorithm.
2. The method for detecting ice coating on a power transmission line based on a wide area perception dynamic convolution model according to claim 1, wherein the step 1 is specifically implemented according to the following steps: Step 1.1, shooting an icing image on a transmission line by using an unmanned aerial vehicle for inspection, and performing horizontal overturning, random matting, rotation transformation, brightness enhancement and contrast enhancement on the obtained icing image of the transmission line to expand a data set; And step 1.2, re-integrating the power transmission line icing image shot in the step 1.1 and the extended power transmission line icing image into a new data set serving as a sample library.
3. The method for detecting ice coating on a power transmission line based on a wide area perception dynamic convolution model according to claim 1, wherein the step 3 is specifically implemented according to the following steps: step 3.1, selecting a power transmission line icing picture to be marked, accurately judging and marking a power transmission line icing region, obtaining a marking frame of the region where the power transmission line icing is positioned, and dividing the power transmission line icing picture into two kinds of labels of the power transmission line icing and the insulator icing according to the power transmission line icing picture in the marking frame, so as to finally obtain a power transmission line icing region marking frame with category labels and an insulator icing region marking frame; Step 3.2, generating XML tag files corresponding to the ice-covered image of the power transmission line by utilizing Labellmg labeling software, storing labeled sample data according to a Pascal VOC format, wherein the sample data file is VOC devkit and comprises three files of Annotations, JPEG Images and IMAGE SETS, and storing all the labeled XML tag files of the ice-covered image and the insulator of the power transmission line in Annotations, wherein the XML tag files comprise image IDs, image paths, image names and pixel heights and widths of Images, the pixel heights and the widths of the Images are represented by four coordinates of a rectangular frame and the XML tag files comprise , , , Wherein% , ) Is the coordinates of the upper left vertex of the rectangular box, (-) -A/D , ) The method comprises the steps of marking the original pictures in JPEG Images, wherein the marked files correspond to the original pictures in the JPEG Images, the JPEG Images comprise all marked original pictures, the Main subfolder is IMAGE SETS, the train document comprises the picture names of all training sets, and the val document comprises the picture names of all verification sets.
4. The method for detecting ice coating on a power transmission line based on a wide area perception dynamic convolution model according to claim 1, wherein the step 4 is specifically implemented according to the following steps: step 4.1, replacing the common convolution in the backbone network with an MDAM multi-dimensional attention module by using the MDAM multi-dimensional attention module, wherein the steps are as follows: the multidimensional attention mechanism comprises 3 branches For the input feature map of the multidimensional attention mechanism, features are extracted by a convolution layer with a ReLU activation function to obtain the feature map Then the characteristic diagram The feature is further extracted by the prior multi-layer perceptron MLP after pooling operation, and the feature map is obtained after multiplying the feature further extracted by a Sigmoid activation function and a channel statistical coefficient Finally, the characteristic diagram is further displayed Input feature map with multidimensional attention mechanism Element-by-element addition to obtain outputs of each branch ; The feature map calculation is shown in the following formula 1: 1 (1) Wherein: representing a first step of feature extraction operation, including two-layer convolution operation and one ReLU activation operation; An input feature map representing a multi-dimensional attention mechanism; Represent the first A feature map obtained after the branch primary feature extraction is completed, ; For characteristic diagram The further extraction of the characteristics is shown in the following steps of formula 2, formula 3 and formula 4: 2, 2 3 4. The method is to Wherein: is an execution operation of a multidimensional attention mechanism; Representing a multi-dimensional statistical coefficient; represents the first Features initially extracted by branch multidimensional attention mechanism pass through activation function and feature As a result of the multiplication the result of which, Taking 1,2 and 3; representing an average pooling operation; representing a maximum pooling operation; Each output of the final multi-dimensional attention module The expression is shown in formula 5: 5. The method is to Wherein: representing the first of the multidimensional attention mechanisms Branching a final output feature map; The multidimensional attention mechanism has three branches with the same structure and different convolution kernel sizes, wherein Conv1, conv2 and Conv3 convolution kernel sizes are respectively , , Therefore, the multidimensional attention mechanism extracts characteristic information from three different scales respectively to obtain a characteristic diagram 、、 And use Is fused with feature maps by convolution and upsampling operations 、、 As shown in formula 6: 6. The method is to Wherein: Representing the operation of the fusion, Representing the operation of the connection and, An output feature map representing a multi-dimensional attention mechanism; and 4.2, introducing a mixed convolution RS Conv into the Thin-head feature extraction network fused with the light model, wherein the mixed convolution RS Conv comprises a shuffling operation and a DW Conv, and the Thin-head feature extraction network fused with the light model is specifically as follows: The RS Conv uses a learnable weight matrix to calculate the weighted sum of the neighbor node characteristics of the nodes, splices the weighted sum with the node characteristics of the RS Conv, and then processes the spliced characteristics through a nonlinear activation function to obtain updated node representation; the method comprises the steps of replacing standardized convolution with RS Conv in an encoder head, introducing RS-bottleneck and VoV-RSCSPC modules into the Thin-head based on the RS Conv, and replacing RepC modules in a baseline model with VoV-RSCSPC modules; Step 4.3, solving the problem that the sample cannot be converged rapidly in the training process by using the Shape-IoU loss function, and simultaneously solving the problems that the baseline model GIoU ignores the similarity of the image pixel level and the global structure and the distribution condition of the image by combining the normalization strategy, wherein the Shape-NWD loss function has the following calculation formula: The formula of Shape-IoU is shown below: 7. The method of the invention 8. The method is used for preparing the product 9. The invention is applicable to 10. The method of the invention 11. The method of the invention 12. Fig. Wherein: 、 For the width and height of the bounding box, 、 The width and the height of the actual frame; And Representing the width and height of the minimum detection frame covering the GT frame and the Anchor frame, respectively, scale being a scaling factor related to the size of the object in the dataset, ww and hh representing weight coefficients in the horizontal direction and in the vertical direction, respectively, the values of which are related to the shape of the GT frame; The corresponding bounding box shape loss is as follows: 13 of the group The idea of Shape-IoU is integrated into the normalized NWD to obtain Shape-NWD, the formula is as follows: 14, of the order of magnitude 15 Of the formula 16, Respectively Wherein D is Euclidean distance between the center point of the GT box and the center point of the anchor frame, =2, C is a constant related to the dataset; And 4.4, forming MDAL-DETR target detection network by the MDAM in the step 4.1, the Thin-head characteristic extraction network fused with the light model in the step 4.2 and the Shape-NWD loss function in the step 4.3.
5. The method for detecting ice coating on a power transmission line based on a wide area sensing dynamic convolution model according to claim 4, wherein the step 5 is specifically implemented according to the following steps: Step 5.1, inputting the training set images divided in the step 2 into the MDAL-DETR target detection network obtained in the step 4.4 for training, and finally obtaining an optimized wide area perception dynamic convolution model to obtain optimal weight data; Step 5.2, two files, mdal-detr. Py and predict. Py, are needed for training result prediction, namely, mdal-detr. Py, model_path and classes _path are needed to be modified first, wherein the model_path points to a trained optimal weight file, classes _path points to txt corresponding to a detection category in a logs folder, and prediction is started after modification is completed.
6. The method for detecting ice coating on a power transmission line based on a wide area perception dynamic convolution model according to claim 5, wherein the specific method for inputting the training set image divided in step 2 into the MDAL-DETR target detection network obtained in step 4.4 for training in step 5.1 is as follows: Setting random gradient reduction with momentum of 0.9 in a file train.py, training 250 rounds, setting input image pixel size 640 x 640, freezing training 50 rounds of batch_size32, thawing training 200 rounds of batch_size to be 4, num_works 2, adam optimizer, attenuation weight coefficient 5 x 10-4, initial learning rate 1 x 10-5, setting IoU threshold value to be 0.5 for test when training set test, fine-tuning learning rate to be 0.003 for better robustness when training, freezing training the first 50 rounds, loss reduction speed fast, thawing training the second 200 rounds, continuously fine-tuning the network, gradually reducing loss change of training set after 200 rounds in total 250 rounds, finally obtaining optimized wide-area perception dynamic convolution model, and obtaining optimal weight data.

Description

Power transmission line icing detection method based on wide-area perception dynamic convolution model Technical Field The invention belongs to the technical field of image processing, and particularly relates to a power transmission line icing detection method based on a wide-area perception dynamic convolution model. Background The transmission line has certain specificity, is often distributed in different areas, has obvious dispersibility, is influenced by various factors such as topography, weather conditions and the like, and has great difficulty in maintenance and management. Moreover, the high-voltage transmission system has great influence on politics and economy due to large capacity, long-distance transmission and wide radiation area, so that the reliability of the high-voltage transmission system needs to reach extremely high standards. However, these lines often span severe environments such as high altitudes, ice coating, acid rain, etc., where the ice coating problem is unavoidable and may lead to reduced power transmission line performance and ice coating accidents. The insulator is a device which is arranged between conductors with different electric potentials or between the conductors and a grounding member and can withstand the actions of voltage and mechanical stress, is a special insulating control, and is widely applied to various electric power systems, including overhead transmission lines, power plants, power substations and the like. Their main function is to support and fix the wires while achieving electrical insulation, ensuring safe and stable operation of the power system. The insulator can be exposed in a bad natural environment for a long time to cause the faults of pollution, flashover, damage, self-explosion, cap shortage and the like of the insulator, so that a great hidden trouble is buried in a safe and stable working environment of electric power. Therefore, whether the problems in the insulator can be found in advance or not, so that the reliable and safe electric energy is ensured, and the primary task of people is realized. The ice coating of the power transmission line and the insulator forms a great risk for the power system, and the weight of the ice can lead to line breakage, tower collapse and insulator breakage. Disclosure of Invention The invention aims to provide a power transmission line icing detection method based on a wide-area perception dynamic convolution model, which solves the defect that the detection performance of a target in a complex environment is not considered in the prior art, The technical scheme adopted by the invention is that the method for detecting the icing of the power transmission line based on the wide-area perception dynamic convolution model is implemented according to the following steps: Step 1, acquiring ice-covered images of the power transmission line in different environments through unmanned aerial vehicle inspection, expanding the obtained ice-covered images of the power transmission line, and taking the expanded ice-covered images of the power transmission line as a real insulator data set; step 2, dividing the ice coating data set of the power transmission line obtained in the step 1 into a training set and a verification set, wherein the ratio of the training set to the verification set is 9:1; step 3, labeling the training set in the step 2 for the ice coating type of the power transmission line by utilizing Labellmg software, labeling the training set by using a VOC format in image labeling, and storing the image information of the ice coating of the power transmission line after labeling to obtain an ice coating database of the power transmission line; step 4, replacing the common convolution in the backbone network with an MDAM multidimensional attention module, wherein the MDAM extracts characteristic information from three different scales to obtain a characteristic map, and fusing the characteristic map by using convolution and up-sampling operation; The method comprises the steps of merging a Thin-head feature extraction network of a lightweight model, introducing a mixed convolution RS Conv, including shuffling operation and DW Conv, calculating a weighted sum of neighbor node features of nodes by the RS Conv by using a learnable weight matrix, splicing the weighted sum with own node features, and then processing the spliced features by a nonlinear activation function to obtain updated node representation; the idea of Shape-IoU is fused into the normalized NWD to obtain Shape-NWD; the MDAM, the Thin-head feature extraction network fused with the light model and the Shape-NWD loss function are combined to form a MDAL-DETR target detection network; Step 5, predicting the verification set in the step 2 by utilizing the optimal weight network obtained after training in the step 4 to obtain a corresponding power transmission line icing detection result; And 6, evaluating the detection result obtained in the step