CN-118711056-B - Transmission line construction equipment sensing method and system based on improved YOLOv model

CN118711056BCN 118711056 BCN118711056 BCN 118711056BCN-118711056-B

Abstract

A transmission line construction equipment perception method and system based on an improved YOLOv model, the method firstly trains an improved YOLOv model which is constructed by a neck network with a cross-connection structure and a bidirectional weighted fusion characteristic pyramid network based on historical transmission line construction equipment images, and then inputs the transmission line construction equipment images shot in real time into the trained improved YOLOv model to obtain a perception result of the transmission line construction equipment. The invention effectively reduces the interference of noise on the model precision and improves the convergence speed and the robustness of the model.

Inventors

CHEN RAN
SUN LIPING
LIAO XIAOHONG
XIONG CHUANYU
LI ZHIWEI
MA LI
QIAO SHIHUI
XIONG YI
ZHANG ZHAOYANG
Shu Sirui
XU CUIPING
XU HAOTIAN
LEI LEI
ZHOU LI
ZHANG HONG
XU HANPING
CAI JIE
HE LANFEI
Li Lvman
ZHOU YINGBO

Assignees

国网湖北省电力有限公司经济技术研究院
湖北科能电力电子有限公司

Dates

Publication Date: 20260508
Application Date: 20240612

Claims (8)

1. A transmission line construction equipment sensing method based on an improved YOLOv model is characterized in that: Comprising the following steps: S1, training a constructed improved YOLOv model based on historical power transmission line construction equipment images, wherein the training comprises the following steps: extracting features of the power transmission line construction equipment image by improving an MP module in a YOLOv7 model neck network, carrying out feature fusion on the extracted power transmission line construction equipment image features by the neck network, updating network weights by using gradients of a loss function, and continuously carrying out iterative optimization until an optimal super parameter is obtained; the neck network is a characteristic pyramid network with a cross-connection structure and two-way weighting fusion; The MP module adopts a channel priori attention mechanism module improved by a high-efficiency channel attention mechanism module, and the module finally outputs a characteristic diagram The method comprises the following steps: ; ; In the above-mentioned method, the step of, Output feature diagram for high-efficiency channel attention mechanism module The channel prior attention mechanism module outputs a feature map for the input, In the case of a deep convolution, Is the ith branch, and For the residual connection, Is the convolution kernel size is Is a convolution operation of (1); s2, inputting the real-time photographed image of the power transmission line construction equipment into a trained improved YOLOv model to obtain a perception result of the power transmission line construction equipment.
2. The transmission line construction equipment sensing method based on the improved YOLOv model according to claim 1, wherein the method comprises the following steps: the cross-connection structure comprises a second CBS module of a convolution layer in a neck network and a first Concat module of a high layer; The bidirectional weighted fusion comprises top-down weighted fusion and bottom-up weighted fusion, each feature fusion is followed by a separable convolution layer, and a specific fusion weight distribution strategy is adopted, wherein the weighted fusion process is as follows: ; ; ; ; In the above-mentioned method, the step of, 、 Input and output characteristics of the i-th layer, As an intermediate property of the i-th layer, , For a depth-separable convolution operation, The fusion weight of the j-th fusion node, Is a minimum value; The specific fusion weight distribution strategy comprises the steps that each characteristic fusion node is endowed with an initial fusion weight value, when the network performs back propagation, the gradient of each fusion weight can be calculated based on a loss function, and then the value of the fusion weight is updated according to a gradient descent method and iterated continuously until the maximum iteration times is reached, wherein the updating formula of the fusion weight is as follows: ; In the above-mentioned method, the step of, 、 Respectively the first The fusion nodes update the fusion weights before and after the updating, In order for the rate of learning to be high, Gradient of the fusion weight for the loss function.
3. The transmission line construction equipment sensing method based on the improved YOLOv model according to claim 1 or 2, wherein the method comprises the following steps: The loss function is EIOU loss function based on Huber loss improvement, and the expression is: ; ; In the above-mentioned method, the step of, As a function of the loss, Based on the cross-over ratio Is a loss of Huber of (a) and, To predict the euclidean distance of the frame center point from the actual frame center point, To predict the euclidean distance of the frame width from the actual frame width, The Euclidean distance between the predicted frame height and the actual frame height, 、 The width and the height of the minimum circumscribed rectangular frame are respectively, To control the transition of the loss function from mean square error to a threshold value of mean absolute error.
4. The transmission line construction equipment sensing method based on the improved YOLOv model according to claim 1, wherein the method comprises the following steps: The output characteristic diagram The method is based on the following steps: Input device Feature map of dimensions Global average pooling is carried out to obtain the product with the size of The obtained characteristic diagram is subjected to convolution kernel with the size of 1-Dimensional convolution and Sigmoid function of (2), and carrying out weight normalization to obtain a size of Multiplying the obtained feature map by the input feature map to obtain an output feature map Wherein the weight is Convolution kernel size Calculated according to the following formula: ; ; In the above-mentioned method, the step of, For the sigmoid activation function, In the form of a one-dimensional convolution, As a feature of the channel(s), As the number of channels of the feature map, 、 Is the parameter of the ultrasonic wave to be used as the ultrasonic wave, Is of an odd nature and is used for screening out variables with odd absolute values.
5. A transmission line construction equipment perception system based on improve YOLOv model, its characterized in that: The power transmission line construction equipment perception module comprises an improved YOLOv model construction module, an improved YOLOv model training module and a power transmission line construction equipment perception module; the improvement YOLOv model building module is configured to build an improvement YOLOv model, wherein a neck network of the improvement YOLOv model is a feature pyramid network with a cross-connection structure and a bi-directional weighted fusion; the improved YOLOv model training module is used for training the constructed improved YOLOv model based on historical transmission line construction equipment images, and comprises the following steps: Extracting features from the power transmission line construction equipment image based on an MP module in a neck network of the improved YOLOv model, carrying out feature fusion on the extracted power transmission line construction equipment image features through the neck network, updating network weights by using gradients of a loss function, and continuously carrying out iterative optimization until an optimal super-parameter is obtained; The MP module adopts a channel priori attention mechanism module improved by a high-efficiency channel attention mechanism module, and the module finally outputs a characteristic diagram The method comprises the following steps: ; ; In the above-mentioned method, the step of, Output feature diagram for high-efficiency channel attention mechanism module The channel prior attention mechanism module outputs a feature map for the input, In the case of a deep convolution, Is the ith branch, and For the residual connection, Is the convolution kernel size is Is a convolution operation of (1); the transmission line construction equipment sensing module is used for inputting the transmission line construction equipment image shot in real time into a trained improved YOLOv model to obtain a sensing result of the transmission line construction equipment.
6. The transmission line construction equipment perception system based on the improved YOLOv model as claimed in claim 5, wherein: the cross-connection structure comprises a second CBS module of a convolution layer in a neck network and a first Concat module of a high layer; The bidirectional weighted fusion comprises top-down weighted fusion and bottom-up weighted fusion, each feature fusion is followed by a separable convolution layer, and a specific fusion weight distribution strategy is adopted, wherein the weighted fusion process is as follows: ; ; ; ; In the above-mentioned method, the step of, 、 Input and output characteristics of the i-th layer, As an intermediate property of the i-th layer, , For a depth-separable convolution operation, The fusion weight of the j-th fusion node, Is a minimum value; The specific fusion weight distribution strategy comprises the steps that each characteristic fusion node is endowed with an initial fusion weight value, when the network performs back propagation, the gradient of each fusion weight can be calculated based on a loss function, and then the value of the fusion weight is updated according to a gradient descent method and iterated continuously until the maximum iteration times is reached, wherein the updating formula of the fusion weight is as follows: ; In the above-mentioned method, the step of, 、 Respectively the first The fusion nodes update the fusion weights before and after the updating, In order for the rate of learning to be high, Gradient of the fusion weight for the loss function.
7. The transmission line construction equipment perception system based on the improved YOLOv model according to claim 5 or 6, wherein: The loss function is EIOU loss function based on Huber loss improvement, and the expression is: ; ; In the above-mentioned method, the step of, As a function of the loss, Based on the cross-over ratio Is a loss of Huber of (a) and, To predict the euclidean distance of the frame center point from the actual frame center point, To predict the euclidean distance of the frame width from the actual frame width, The Euclidean distance between the predicted frame height and the actual frame height, 、 The width and the height of the minimum circumscribed rectangular frame are respectively, To control the transition of the loss function from mean square error to a threshold value of mean absolute error.
8. The transmission line construction equipment perception system based on the improved YOLOv model as claimed in claim 5, wherein: The output characteristic diagram The method is based on the following steps: Input device Feature map of dimensions Global average pooling is carried out to obtain the product with the size of The obtained characteristic diagram is subjected to convolution kernel with the size of 1-Dimensional convolution and Sigmoid function of (2), and carrying out weight normalization to obtain a size of Multiplying the obtained feature map by the input feature map to obtain an output feature map Wherein the weight is Convolution kernel size Calculated according to the following formula: ; ; In the above-mentioned method, the step of, For the sigmoid activation function, In the form of a one-dimensional convolution, As a feature of the channel(s), As the number of channels of the feature map, 、 Is the parameter of the ultrasonic wave to be used as the ultrasonic wave, Is of an odd nature and is used for screening out variables with odd absolute values.

Description

Transmission line construction equipment sensing method and system based on improved YOLOv model Technical Field The invention belongs to the technical field of power transmission line detection, and particularly relates to a power transmission line construction equipment sensing method and system based on YOLOv model. Background Along with the rapid development of smart power grids, the power transmission line is taken as an important component of a power system, the level of intellectualization of construction and maintenance is also continuously improved, and literature reports for detecting related equipment of the power transmission line based on machine learning are also increasing. For example, in the literature ' transmission line aerial insulator detection based on improvement FASTER RCNN ' (Yi Jiyu, chen Cifa, guojiang [ J ]. Computer engineering, 2021,47 (06): 292-298+304.) ' a difficult sample is generated by introducing a shielding mask to perform training and learning, so that the insulator detection precision of a FASTER RCNN model under shielding condition is improved. However, the two-stage detection algorithm FASTER RCNN requires that candidate regions be generated first, then classified and located, which is much slower than the one-stage algorithm in detection speed. The wind turbine blade surface defect detection algorithm [ J ]. Chinese electric power, 2023,56 (10): 43-52 ] based on HSCA-YOLOv is introduced into an improved space pyramid pooling module on the basis of a YOLOv model, a mixed space channel attention mechanism is provided, and the accuracy of detecting the surface defects of the wind turbine blade by the model is improved. However, in practical situations, the YOLOv model still has the problem of losing characteristic information, and the accuracy of small target detection still needs to be improved. Disclosure of Invention The invention aims to solve the problems in the prior art and provides a transmission line construction equipment sensing method and system based on YOLOv model. In order to achieve the above object, the technical scheme of the present invention is as follows: In a first aspect, the present invention provides a transmission line construction equipment sensing method based on an improved YOLOv model, including: S1, training a constructed improved YOLOv model based on historical power transmission line construction equipment images, wherein the training comprises the following steps: performing feature fusion on the extracted image features of the power transmission line construction equipment through a neck network of the improved YOLOv7 model, updating network weights by using gradients of a loss function, and continuously iterating and optimizing until an optimal super-parameter is obtained; the neck network is a characteristic pyramid network with a cross-connection structure and two-way weighting fusion; s2, inputting the real-time photographed image of the power transmission line construction equipment into a trained improved YOLOv model to obtain a perception result of the power transmission line construction equipment. The cross-connection structure comprises a second CBS module of a convolution layer in a neck network and a first Concat module of a high layer; The bidirectional weighted fusion comprises top-down weighted fusion and bottom-up weighted fusion, each feature fusion is followed by a separable convolution layer, and a specific fusion weight distribution strategy is adopted, wherein the weighted fusion process is as follows: In the above-mentioned method, the step of, P iout is the input and output characteristics of the ith layer, P itd is the middle characteristic of the ith layer, i=3, 4,5, D (·) is the depth separable convolution operation, ω j is the fusion weight of the jth fusion node, and ε is a minimum value; The specific fusion weight distribution strategy comprises the steps that each characteristic fusion node is endowed with an initial fusion weight value, when the network performs back propagation, the gradient of each fusion weight can be calculated based on a loss function, and then the value of the fusion weight is updated according to a gradient descent method and iterated continuously until the maximum iteration times is reached, wherein the updating formula of the fusion weight is as follows: In the above formula, omega l、ωl' is the fusion weight before and after updating of the first fusion node, In order for the rate of learning to be high,Gradient of the fusion weight for the loss function. The loss function is EIOU loss function based on Huber loss improvement, and the expression is: In the above formula, L Huber-EIoU is a loss function, L δ (IoU, 1) is Huber loss based on an intersection ratio IoU, ρ (b, b gt) is a euclidean distance between a predicted frame center point and an actual frame center point, ρ (w, w gt) is a euclidean distance between a predicted frame width and an actual frame width, ρ (h, h gt)