CN-122020478-A - Power transmission line hidden danger detection method based on model light weight

CN122020478ACN 122020478 ACN122020478 ACN 122020478ACN-122020478-A

Abstract

The invention relates to the technical field of intelligent monitoring of power transmission lines, in particular to a method for detecting hidden danger of a power transmission line based on model weight, which comprises the steps of obtaining an image and a sensing data stream of the power transmission line, performing pattern deconstructment on the data stream according to a predefined rule, separating an abnormal pattern and a background pattern and forming heterogeneous data pairs. The heterogeneous data pair is input into a lightweight detection network, and through parallel branches of the lightweight detection network, the dynamic convolution check abnormal data are adopted to conduct feature extraction to generate an abnormal feature map, and the static convolution check background data are adopted to conduct environment modeling to generate an environment reference map. And calculating the relevance of the two maps by the fusion judgment branch, and outputting a structural hidden danger report according to the relevance. According to the method, through data mode decoupling and heterogeneous parallel processing, the detection precision is kept, meanwhile, the model calculation complexity and the resource consumption are reduced, and real-time and accurate hidden danger identification on the edge side with limited resources is realized.

Inventors

LI XIN
ZHAO LIANGUO
DU YANG
SHI FENG
LIN SEN
CHEN LINGLING

Assignees

国网山东省电力公司费县供电公司

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (10)

1. The utility model relates to a method for detecting hidden danger of a power transmission line based on model light weight, which is characterized by comprising the following steps: Acquiring an original monitoring data stream of the power transmission line, wherein the original monitoring data stream comprises image data and time sequence sensing data; Performing mode deconstructment on the original monitoring data stream of the power transmission line according to a predefined hidden danger characterization rule, and separating an abnormal mode data stream and a background mode data stream, wherein the abnormal mode data stream and the background mode data stream form heterogeneous data pairs; Inputting the heterogeneous data pairs into a preloaded lightweight detection network, wherein the lightweight detection network comprises a parallel processing branch and a fusion judgment branch; In the parallel processing branch, dynamic convolution is adopted to check the abnormal pattern data flow to conduct feature extraction, and static convolution is adopted to check the background pattern data flow to conduct environment modeling at the same time, so that an abnormal feature map and an environment reference map are respectively generated; Inputting the abnormal characteristic spectrum and the environment reference spectrum into the fusion judgment branch, and executing cross-spectrum relevance calculation; And generating a structural hidden danger report containing hidden danger positions and hidden danger types according to the correlation calculation result.
2. The method for detecting hidden danger of a power transmission line based on model weight reduction according to claim 1, wherein the step of performing mode deconstructing on the power transmission line original monitoring data stream according to a predefined hidden danger characterization rule specifically comprises: Invoking an edge significance detection operator aiming at image data in the original monitoring data stream of the power transmission line, and extracting a high gradient change region and a low texture smooth region in an image; Aiming at time sequence sensing data in the original monitoring data stream of the power transmission line, a sliding window differential algorithm is applied to identify mutation points and periodic fluctuation segments in a data sequence; performing space-time alignment and logic binding on the high gradient change region and the mutation points to combine the high gradient change region and the mutation points into an initial candidate set of the abnormal mode data stream; Performing data fusion on the low-texture smooth region and the periodic fluctuation segment to form a stable reference set of the background mode data stream; And applying confidence filtering to the initial candidate set of the abnormal mode data stream, and screening out the pseudo abnormal data with the confidence lower than a preset threshold value to form a final abnormal mode data stream.
3. The method for detecting hidden danger of a power transmission line based on model weight according to claim 2, further comprising, before the step of applying confidence filtering to the initial candidate set of the abnormal pattern data stream: constructing a multi-level confidence coefficient model for evaluating the pseudo-abnormal data, wherein the multi-level confidence coefficient model integrates historical false alarm statistical characteristics and real-time environment interference factors; Extracting space-time attribute and physical quantity attribute item by item from the initial candidate set of the abnormal mode data stream; inputting the space-time attribute and the physical quantity attribute of each item into the multi-level confidence coefficient model, and calculating to obtain a single confidence coefficient score; generating a confidence distribution histogram of the initial candidate set of the abnormal mode data stream according to all the single confidence scores; And dynamically adjusting the numerical value of the preset threshold based on the confidence distribution histogram.
4. The method for detecting hidden danger of a power transmission line based on model weight reduction according to claim 1, wherein in the step of inputting the heterogeneous data pair into a preloaded lightweight detection network, a loading process of the lightweight detection network includes: Downloading a basic network topology structure matched with the current transmission line tower model from a cloud model warehouse; reading a hardware performance configuration file deployed on local edge equipment, wherein the hardware performance configuration file comprises memory capacity and processor computing power information; performing online pruning operation on the basic network topological structure according to the hardware performance configuration file, and removing redundant connection layers and activation layers; Using the network topology structure generated after pruning to incrementally pull the corresponding model weight parameter file from the cloud model warehouse; and loading the model weight parameter file into a memory to complete the instantiation and the preheating of the lightweight detection network.
5. The method for detecting hidden danger of a power transmission line based on model weight according to claim 4, wherein the step of performing an online pruning operation on the basic network topology specifically comprises: analyzing contribution degree weight of each layer of network in the basic network topology structure to a history hidden danger detection task; according to the contribution degree weight, sorting the network layers, and marking a network layer set to be removed, the contribution degree of which is lower than a pruning threshold value; Checking the data dependency relationship between the network layer in the network layer set to be removed and the front and back adjacent network layers; For the network layer with strong data dependency, the structure is reserved, the weight is reset to zero, and for the network layer without strong data dependency, the structure is directly deleted from the basic network topology structure; traversing and updating the data flow path of the whole network to generate a simplified network topology description file.
6. The method for detecting hidden danger of a power transmission line based on model weight reduction according to claim 1, wherein the step of feature refinement of the abnormal pattern data stream by dynamic convolution check comprises the following steps: calculating the size parameter and the step size parameter of the convolution kernel in real time according to the statistical characteristics of the data segments in the abnormal mode data stream; Dynamically assembling a convolution calculation unit in a memory by using the calculated size parameter and step size parameter; dividing the abnormal mode data stream into a plurality of data blocks, and sequentially inputting the data blocks into the dynamically assembled convolution calculation unit for convolution operation; Collecting an intermediate feature map output by each data block after convolution operation; And carrying out multi-scale pooling and normalization treatment on all the intermediate feature maps, and integrating to form the abnormal feature map.
7. The method for detecting hidden danger of a power transmission line based on model weight reduction according to claim 1, wherein the step of performing cross-map relevance calculation specifically comprises: gridding and aligning the abnormal characteristic spectrum and the environment reference spectrum to ensure that the abnormal characteristic spectrum and the environment reference spectrum have grid units in one-to-one correspondence on space coordinates; Extracting feature vectors from the abnormal feature map and the environment reference map for each aligned grid cell; calculating cosine similarity and Euclidean distance between the feature vector from the abnormal feature map and the feature vector from the environment reference map; combining the cosine similarity and the Euclidean distance into a relevance strength index; Traversing all grid units to generate a correlation intensity distribution map with the same spatial resolution as the original map.
8. The method for detecting hidden danger of a power transmission line based on model weight reduction according to claim 7, wherein after generating a correlation intensity distribution map with the same spatial resolution as the original map, further comprises: Performing region growing segmentation on the correlation intensity distribution map, and identifying a communication region with correlation intensity exceeding a judgment threshold value; calculating the geometric center coordinates and the circumscribed rectangle boundary of each communication area; backtracking to the corresponding position of the original abnormal feature map according to the geometric center coordinates, and extracting a multi-mode sensing data original fragment corresponding to the position; And combining the circumscribed rectangular boundary with the extracted multi-mode sensing data original fragment to generate a description field about hidden danger positions and a classification label of hidden danger types in the structural hidden danger report.
9. The method for detecting hidden danger of a power transmission line based on model weight according to claim 1, wherein the step of obtaining the original monitoring data stream of the power transmission line including the image data and the time sequence sensing data specifically comprises: Polling data acquisition proxy service deployed at each monitoring point of the power transmission line; When the data acquisition proxy service returns to a ready state, sending a data pulling instruction with a time stamp to the data acquisition proxy service; receiving a packed data block returned by the data acquisition proxy service according to the data pulling instruction; unpacking and checking the packed data blocks to separate independent image data files and time sequence sensing data files; And synchronizing and splicing the image data file and the time sequence sensing data file according to a unified time reference to form the original monitoring data stream of the power transmission line.
10. The method for detecting hidden danger of a power transmission line based on model weight reduction according to claim 9, wherein after the step of unpacking and verifying the packed data block, further comprising: If the verification fails, recording the equipment identification and failure time of the data acquisition proxy service corresponding to the packed data block; transmitting a data retransmission request to the data acquisition proxy service, and starting a retransmission timer; If the correct data packet is received before the retransmission timer is overtime, continuing the processing flow; And if the retransmission timer is overtime and still does not receive the correct data packet, marking the equipment identification as abnormal, temporarily removing the equipment identification from the current polling list, and simultaneously enabling the standby data acquisition proxy service of the same area to carry out data supplement acquisition.

Description

Power transmission line hidden danger detection method based on model light weight Technical Field The invention relates to the technical field of intelligent monitoring of power transmission lines, in particular to a power transmission line hidden danger detection method based on model weight reduction. Background Current intelligent transmission line monitoring relies on analyzing a continuous data stream of image data and time-series sensing data. The prior art generally adopts an end-to-end deep learning model to process such multi-source monitoring data. The models take the original data stream as an integral input, and directly learn and output hidden danger identification results through a complex network structure. The method aims to realize automatic detection of various hidden dangers, and the effectiveness of the method is established on the basis that the model has strong enough feature extraction and fusion capability. Such unified treatments have drawbacks. The monitoring data stream is essentially a mixture of burst anomalies and steady state background information, and prior methods have failed to effectively distinguish between these two components in the process flow. The background and the abnormal information are input into a single model together, network parameters are forced to be fitted with data modes with obvious differences at the same time, and the complexity and convergence difficulty of model training are increased. In order to achieve acceptable detection accuracy, the system often has to rely on models with deeper structures and larger parameter amounts, resulting in high computational complexity and large resource consumption. The existing scheme is difficult to realize efficient deployment and real-time analysis on the edge monitoring terminal with limited computing capacity and storage resources. The detection method capable of optimizing from the data processing source and matching with the light model framework is urgently needed, so that the requirement on hardware calculation force is reduced while the detection precision is ensured, and the actual requirements on real-time performance and low power consumption in a distributed monitoring scene of a power transmission line are met. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a model-based lightweight power transmission line hidden danger detection method. In order to achieve the purpose, the invention adopts the following technical scheme that the method for detecting hidden danger of the power transmission line based on model weight reduction comprises the following steps: Acquiring an original monitoring data stream of the power transmission line, wherein the original monitoring data stream comprises image data and time sequence sensing data; Performing mode deconstructment on the original monitoring data stream of the power transmission line according to a predefined hidden danger characterization rule, and separating an abnormal mode data stream and a background mode data stream, wherein the abnormal mode data stream and the background mode data stream form heterogeneous data pairs; Inputting the heterogeneous data pairs into a preloaded lightweight detection network, wherein the lightweight detection network comprises a parallel processing branch and a fusion judgment branch; In the parallel processing branch, dynamic convolution is adopted to check the abnormal pattern data flow to conduct feature extraction, and static convolution is adopted to check the background pattern data flow to conduct environment modeling at the same time, so that an abnormal feature map and an environment reference map are respectively generated; Inputting the abnormal characteristic spectrum and the environment reference spectrum into the fusion judgment branch, and executing cross-spectrum relevance calculation; And generating a structural hidden danger report containing hidden danger positions and hidden danger types according to the correlation calculation result. As a further aspect of the present invention, the step of performing mode deconstructing on the original monitoring data stream of the power transmission line according to a predefined hidden danger characterization rule specifically includes: Invoking an edge significance detection operator aiming at image data in the original monitoring data stream of the power transmission line, and extracting a high gradient change region and a low texture smooth region in an image; Aiming at time sequence sensing data in the original monitoring data stream of the power transmission line, a sliding window differential algorithm is applied to identify mutation points and periodic fluctuation segments in a data sequence; performing space-time alignment and logic binding on the high gradient change region and the mutation points to combine the high gradient change region and the mutation points into an initial candidate set of the abnorm