CN-121564328-B - Trend perception weighted iteration type power grid bird detection method and system

CN121564328BCN 121564328 BCN121564328 BCN 121564328BCN-121564328-B

Abstract

The invention relates to the technical field of grid bird detection, in particular to a trend perception weighting iterative grid bird detection method and system. The method comprises the steps of deploying a lightweight target detection model on edge equipment to detect in real time and output candidate images, transmitting the candidate images to a cloud end in real time, periodically extracting and transmitting images of undetected birds, identifying bird targets and key power equipment by the aid of the cloud end multi-mode large model, calculating Euclidean distances normalized relative to equipment dimensions, calculating a space risk index by combining a vertical azimuth relation and semantic confidence, judging early warning and marking bad examples according to the space risk index, inputting bad example characteristics into a gradient lifting tree model to predict training weight, and carrying out incremental training and redeployment on loss function weighting by using the training weight and the space risk index. According to the invention, through edge cloud coordination, space risk quantization and trend perception weighting iteration, high-precision detection, low false alarm rate and continuous evolution capability of the model are realized.

Inventors

XU RUIZE
OUYANG WENHUA
LI CHANGDONG
SHU XINYU
YANG HAO
FANG GUANGZONG
ZHANG HUAIZHENG

Assignees

南昌科晨电力试验研究有限公司
国网江西省电力有限公司电力科学研究院

Dates

Publication Date: 20260512
Application Date: 20260122

Claims (9)

1. A trend perception weighted iterative power grid bird detection method is characterized by comprising the following steps: S100, deploying a lightweight target detection model on edge computing equipment, detecting a monitoring image in real time, and outputting candidate images containing birds; S200, transmitting candidate images output by an edge end to a cloud end in real time, and periodically extracting and transmitting images in which birds are not detected to the cloud end; When the bird target and the key power equipment are detected at the same time, the space position information and the semantic confidence of the bird target and the key power equipment are extracted, the Euclidean distance normalized relative to the equipment scale between the bird target and the key power equipment is calculated, and the space risk index is calculated by combining the vertical azimuth relation and the semantic confidence; s400, judging whether to early warn according to the space risk index, and marking a missing report, false report or low confidence sample as a bad case; s500, recording the spatial risk index, error type and trend characteristics of the bad case samples, inputting the characteristics into a gradient lifting tree model, and automatically predicting the training weight of each bad case; And S600, weighting a loss function of the bad sample by using the training weight and the space risk index, performing incremental training on the lightweight target detection model, and redeploying the trained lightweight target detection model to the edge equipment.
2. The trend-aware weighted iterative power grid bird detection method of claim 1, wherein the training process of the lightweight target detection model comprises: constructing a power grid bird image data set containing tower ID metadata, and grouping according to the tower IDs; Randomly dividing the tower groups obtained after grouping into K mutually disjoint subsets; Performing K rounds of cross validation training, wherein one subset is selected as a validation set in each round, and the rest subsets are used as training sets; And saving a model with optimal performance in K-round cross validation as the lightweight target detection model.
3. A trend awareness weighted iterative power grid bird detection method in accordance with claim 1, wherein said identifying bird targets and critical power devices comprises: performing global scanning on the image, and detecting whether the image contains bird features and equipment features; Filtering the image when only one type of feature is included; when both types of features are contained, the birds target collection and the key power equipment collection are extracted.
4. The trend awareness weighted iterative power grid bird detection method of claim 3, wherein each bird target in the set of bird targets comprises a detection frame, geometric center point coordinates and semantic confidence, and each key power device in the set of key power devices comprises a detection frame, geometric center point coordinates and semantic confidence.
5. The trend-aware weighted iterative power grid bird detection method of claim 1, wherein said calculating a spatial risk index combining vertical bearing relationships and semantic confidence comprises: Determining a vertical orientation of the avian target relative to the critical power equipment, setting a vertical penalty factor when the avian target is located above the critical power equipment; constructing a distance attenuation function based on the normalized Euclidean distance; And calculating a spatial risk index by combining the semantic confidence coefficient, the distance decay function and the vertical penalty factor.
6. The trend-aware weighted iterative power grid bird detection method of claim 1, wherein marking a missing report, false report, or low confidence sample as bad comprises: For images which are transmitted to the cloud end and are not detected by birds, marking as a missing report bad case when the cloud end detects bird targets and key power equipment at the same time and the space risk index is higher than a high-risk threshold value; for candidate images output by the edge end, marking as false alarm bad examples when the cloud end detects bird targets and key power equipment simultaneously and the space risk index is lower than a safety threshold value; and for the candidate images output by the edge end, marking as a bad example with low confidence coefficient when the cloud end detects the bird target and the key power equipment simultaneously and the space risk index is between the safety threshold and the high-risk threshold.
7. The trend-aware weighted iterative power grid bird detection method of claim 1, wherein the trend characteristics include cumulative number of bad examples, distribution of bad example types, and trend of spatial risk index variation of the tower in a historical time window.
8. The trend-aware weighted iterative power grid bird detection method of claim 1, wherein weighting the loss function of the bad samples using training weights and spatial risk indices comprises: calculating the classification loss and regression loss of each bad sample; adding the classification loss and the regression loss, and multiplying the training weight and the space risk index of the corresponding sample; The weighted losses for all bad samples are summed as the total loss for incremental training.
9. A trend-aware weighted iterative grid bird detection system, comprising: the edge detection module is used for deploying a lightweight target detection model on the edge computing equipment, detecting the monitoring image in real time and outputting candidate images containing birds; The cooperative transmission module is used for transmitting the candidate images output by the edge end to the cloud end in real time, and periodically extracting and transmitting the images of which the birds are not detected to the cloud end; The risk calculation module is used for globally detecting the received image by the cloud multi-mode large model and identifying the bird target and the key power equipment, extracting the space position information and the semantic confidence coefficient of the bird target and the key power equipment when the bird target and the key power equipment are detected at the same time, calculating the Euclidean distance normalized between the bird target and the key power equipment relative to the equipment scale, and calculating the space risk index by combining the vertical azimuth relation and the semantic confidence coefficient; the bad case marking module is used for judging whether to early warn according to the space risk index and marking a missing report, false report or low confidence sample as a bad case; the weight prediction module is used for recording the spatial risk index, the error type and the trend characteristics of the bad case samples, inputting the characteristics into the gradient lifting tree model, and automatically predicting the training weight of each bad case; and the iteration optimization module is used for weighting the loss function of the bad sample by using the training weight and the space risk index, performing incremental training on the lightweight target detection model, and redeploying the trained lightweight target detection model to the edge equipment.

Description

Trend perception weighted iteration type power grid bird detection method and system Technical Field The invention relates to the technical field of grid bird detection, in particular to a trend perception weighting iterative grid bird detection method and system. Background With the expansion of the construction scale of the power grid and the improvement of the ecological environment, bird activities form a serious threat to the safety of the power grid. According to statistics, in 110kV and above lines of power grid companies, bird damage faults account for 10.4% and are in an ascending trend, the single maintenance cost is as high as tens of thousands yuan, and an efficient intelligent detection scheme is needed. In the prior art, a single lightweight target detection model is usually deployed on front-end equipment, and has obvious defects that firstly, the model is limited by edge calculation force, false detection is serious under complex background, small target or bad weather, secondly, long-range birds and background birds cannot be distinguished, so that a large number of invalid early warning is caused, and finally, the performance of the deployed model is solidified, a new sample on site cannot be learned, the continuous optimization capability is lacked, and the long-term operation reliability is not high. Disclosure of Invention The invention provides a trend perception weighted iteration type power grid bird detection method and a trend perception weighted iteration type power grid bird detection system, which solve the problems of insufficient precision, high false alarm rate and model solidification of the existing power grid bird detection method, and realize high-precision sustainable evolution intelligent detection. In order to achieve the above purpose, the present invention provides the following technical solutions: the invention discloses a trend perception weighting iterative power grid bird detection method, which comprises the following steps: S100, deploying a lightweight target detection model on edge computing equipment, detecting a monitoring image in real time, and outputting candidate images containing birds; S200, transmitting candidate images output by an edge end to a cloud end in real time, and periodically extracting and transmitting images in which birds are not detected to the cloud end; When the bird target and the key power equipment are detected at the same time, the space position information and the semantic confidence of the bird target and the key power equipment are extracted, the Euclidean distance normalized relative to the equipment scale between the bird target and the key power equipment is calculated, and the space risk index is calculated by combining the vertical azimuth relation and the semantic confidence; s400, judging whether to early warn according to the space risk index, and marking a missing report, false report or low confidence sample as a bad case; s500, recording the spatial risk index, error type and trend characteristics of the bad case samples, inputting the characteristics into a gradient lifting tree model, and automatically predicting the training weight of each bad case; And S600, weighting a loss function of the bad sample by using the training weight and the space risk index, performing incremental training on the lightweight target detection model, and redeploying the trained lightweight target detection model to the edge equipment. As a preferable technical solution of the present invention, the training process of the lightweight target detection model includes: constructing a power grid bird image data set containing tower ID metadata, and grouping according to the tower IDs; Randomly dividing the tower groups obtained after grouping into K mutually disjoint subsets; Performing K rounds of cross validation training, wherein one subset is selected as a validation set in each round, and the rest subsets are used as training sets; And saving a model with optimal performance in K-round cross validation as the lightweight target detection model. As a preferred embodiment of the present invention, the identifying of birds and key electric power equipment includes: performing global scanning on the image, and detecting whether the image contains bird features and equipment features; Filtering the image when only one type of feature is included; when both types of features are contained, the birds target collection and the key power equipment collection are extracted. According to the optimal technical scheme, each bird target in the bird target set comprises a detection frame, a geometric center point coordinate and a semantic confidence coefficient, and each key power device in the key power device set comprises the detection frame, the geometric center point coordinate and the semantic confidence coefficient. As a preferred technical solution of the present invention, the calculating the spatial risk index by combining the vertical