Power transmission line risk assessment system and method based on image recognition
By using an image recognition-based transmission line risk assessment system, which combines multi-scale feature decomposition, spectral feature extraction, and topological feature extraction, and employs random forest algorithm and directed edge network, the system solves the problems of multi-source data fusion, feature extraction, and dynamic adaptability in transmission line soil erosion monitoring, and achieves efficient and accurate risk identification and dynamic calibration.
Patent Information
- Application Number
- CN202511003509.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-21
- Publication Date
- 2025-11-07
AI Technical Summary
Existing technologies for monitoring soil erosion in power transmission lines suffer from problems such as insufficient multi-source data fusion, single feature extraction dimensions, lagging dynamic adaptability, and an imbalance between efficiency and accuracy, leading to one-sided risk assessment, delayed early warning, and high costs.
A transmission line risk assessment system based on image recognition is adopted. Through multi-source data acquisition, data preprocessing, feature extraction and risk identification modules, combined with multi-scale feature decomposition, spectral feature extraction, topological feature extraction and feature fusion, a random forest algorithm is used for initial screening and a directed edge network is used for fine assessment to achieve real-time dynamic calibration.
It improves the reliability and accuracy of risk identification, achieves a balance between efficiency and precision, supports large-scale and efficient initial screening and detailed investigation of high-risk areas, dynamically responds to environmental changes, and reduces the cost of manual inspections.
Smart Images
Figure CN120913064A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of image risk identification, and in particular to a power transmission line risk assessment system and method based on image recognition. BACKGROUND
[0002] The monitoring of water and soil erosion along the power transmission line plays a key role in ensuring the safe operation of the power grid. The core goal is to identify disaster risks such as tower foundation exposure and landslides in advance through the analysis of multi-dimensional information such as topographic changes, vegetation coverage status, and meteorological environment, to provide scientific decision-making basis for line operation and maintenance, reduce manual inspection costs, and improve the timeliness of disaster warning.
[0003] However, the prior art still has the following significant deficiencies: insufficient multi-source data fusion: traditional monitoring methods cannot effectively integrate remote sensing images, account data, historical maintenance records, real-time meteorological factors, and other types of data, resulting in risk assessment relying on a single data source, which cannot build a comprehensive water and soil erosion influencing factor model, and is prone to one-sided evaluation problems; single feature extraction dimension: only spectral features or geometric features are analyzed, lacking quantitative description of topological properties such as topographic structure connectivity and hole features, making it difficult to capture potential risks of topographic "connectivity-disruption", and easily missing water and soil erosion hazards caused by poor topographic structure stability; dynamic adaptability lag: unable to respond in real time to sudden changes in environmental factors such as heavy rain and seasonal changes in vegetation, risk level assessment relies on fixed thresholds or historical data, lacking a dynamic calibration mechanism based on real-time data, resulting in delayed warning relative to actual risk evolution; imbalance between efficiency and accuracy: traditional physical models can achieve fine-grained calculations, but when faced with large-scale power transmission line monitoring, the calculation complexity is high and time-consuming; manual screening is costly and difficult to balance between "wide-range coarse screening" and "high-risk area detailed inspection", restricting the improvement of monitoring efficiency. SUMMARY
[0004] The technical problem to be solved by the present application is to provide a risk identification system and method that is more timely, efficient, accurate, and reliable.
[0005] To solve the above technical problems, the technical solution adopted by the present application is:
[0006] A power transmission line risk assessment system based on image recognition, the key of which is that the system includes a multi-source data acquisition module, a data preprocessing module, a feature extraction module, and a risk identification module.
[0007] The multi-source data collection module collects external image information, account data, historical maintenance records and real-time meteorological factors, and transmits the image data to the data and processing module for preprocessing, transmits the preprocessed image data to the feature extraction module to obtain a fusion feature map of the image, and finally the risk identification module judges the water and soil erosion risk level of the image through the fusion feature map.
[0008] Preferably, the data preprocessing module comprises a space-time registration module, a noise suppression module and a feature enhancement module.
[0009] The space-time registration module receives the image information of the multi-source data collection module, and transmits the registered image to the noise suppression module, and the de-noised image is sharpened by the feature enhancement module and outputs a sharpened image.
[0010] Preferably, the feature extraction module comprises a multi-scale feature decomposition module, a spectral feature extraction module, a topological feature extraction module and a feature fusion module.
[0011] The multi-scale feature decomposition module receives the sharpened image for feature decomposition to obtain high-frequency detail features and low-frequency approximation features; the spectral feature extraction module receives the sharpened image and outputs a spectral feature vector; the topological feature extraction module receives the low-frequency approximation feature map and outputs a topological feature vector; and the feature fusion module splices the high-frequency detail features, the low-frequency approximation features, the spectral feature vector and the topological feature vector to obtain a fusion feature map.
[0012] A power transmission line risk assessment method based on image recognition, which is characterized by comprising the following steps:
[0013] In the data acquisition stage, image information is obtained by manual shooting, fixed camera shooting and unmanned aerial vehicle shooting, and account data, historical maintenance records and real-time meteorological factors are received through an interface;
[0014] In the data preprocessing stage, the image information and non-image information obtained by the data acquisition module are first aligned according to the time stamp, and the coordinate system is unified through geometric correction; then the registered image is windowed and de-noised, and the de-noised image is processed by the feature enhancement module through a second-order differential sharpening algorithm to obtain a sharpened image;
[0015] In the feature extraction stage, the sharpened image is decomposed into high-frequency detail features reflecting edges and small topography and low-frequency approximation features reflecting macroscopic terrain trends through wavelet transform, and normalized vegetation index and soil brightness index are calculated to obtain NDVI value and SBI value; further, a scale space is constructed for the low-frequency approximation features, persistent pairs are extracted and vectorized into a topological feature vector; finally, the multi-modal features are spliced to obtain a fusion feature map;
[0016] In the risk identification stage, firstly, screening is performed according to the fused feature map, the occurrence probability of water and soil loss is calculated through a random forest algorithm, and a high-risk area is marked; the soil loss amount is further calculated for the high-risk area, and risk grade discrimination is realized through fused feature data.
[0017] Further, in the risk identification stage, the specific process of realizing risk grade discrimination through fused feature data is as follows:
[0018] The directed edge is constructed, the correlation of each input data is performed, and the directed edge node and direction are determined; the directed edge finally points to the risk degree node
[0019] The conditional probability table is constructed, firstly, the continuous quantity in the input data is discretized, and then the conditional probability between nodes is defined through historical data;
[0020] Inference calculation, according to real-time data, the posterior probability of each node is updated, and the posterior probability of the output node is calculated according to the probability of the intermediate node.
[0021] The beneficial effects produced by the above technical scheme are as follows:
[0022] The present application combines spatial, spectral and topological features, realizes comprehensive feature description from micro to macro, improves the richness of feature expression, and further improves the reliability and accuracy of risk identification.
[0023] The present application realizes large-scale preliminary screening through a random forest algorithm, and then finely evaluates the high-risk area through a directed edge network, realizes the balance and combination of efficiency and precision, improves the identification efficiency, and at the same time, as much as possible, preserves the identification accuracy.
[0024] The present application updates the node state in the directed edge network through real-time data, can quickly respond to the change of external environment, realizes dynamic calibration and instant reporting of risk. BRIEF DESCRIPTION OF DRAWINGS
[0025] The present application will be further described in detail below in combination with the drawings and specific embodiments.
[0026] Figure 1 It is a structure schematic diagram of a power transmission line risk assessment system based on image recognition proposed by the present application. DETAILED DESCRIPTION
[0027] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings and specific embodiments in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0028] An image recognition-based power transmission line risk assessment system, comprising the system comprises a multi-source data acquisition module, a data preprocessing module, a feature extraction module, a risk identification module;
[0029] The multi-source data acquisition module collects remote sensing images along the power transmission line, including manual shooting, fixed camera shooting and unmanned aerial vehicle shooting, supports synchronous acquisition of multiple devices and multiple types of data, and integrates timestamp and geographic location metadata for easy data tracing. Through the interface, account data, historical maintenance records and real-time meteorological factors are received and stored.
[0030] The data preprocessing module includes three sub-modules, including a space-time registration module, a noise suppression module and a feature enhancement module.
[0031] The space-time registration module: aligns multi-source images and other information based on timestamps, matches images with meteorological factors and other information, and unifies the coordinate system through geometric correction to solve the space-time misalignment problem and prevent space-time misalignment of multi-source image data.
[0032] Noise suppression module: removes image noise through windowed median filtering to improve the reliability of subsequent image recognition. The median filtering method directly processes the gray value of the pixel point to achieve image denoising. By determining the size of the field template, usually 3x3, 5x5, 7x7 templates are selected. After arranging the gray values of each pixel point in the field in order from small to large, the gray value in the middle of the order is used to replace the gray value of the middle pixel point in the field. Finally, all pixel points in the image are processed in the above manner to obtain the median filtered image g(x,y). The specific formula is:
[0033] g(x,y)=Med{f(x-k,y-l),(k,l∈W)}
[0034] Where W is the size of the field window, and f(x-k,y-l) is the pixel value in the field template. In order to find the center value more simply, the total number of pixel points contained in the W field is generally an odd value.
[0035] Feature enhancement module: uses a second-order differential sharpening algorithm to highlight feature edges and enhance image details. In specific use, the Laplace operator is used, where the Laplace transform of the binary function f(x,y) is defined as:
[0036]
[0037] The second order partial derivatives in x and y directions are:
[0038]
[0039] The above formulas can be integrated to obtain:
[0040]
[0041] The template based on the Laplace operator is obtained, the Laplace operator emphasizes the mutation of the gray scale in the image, and does not emphasize the slow changing area of the image, so some gradually changing light gray edge lines will become the background color of the picture contour, then the original image and the Laplace image are superimposed, the original image is maintained while the enhanced edge image can be seen, thereby realizing the sharpening processing of the image.
[0042] The feature extraction module includes four sub-modules, including a multi-scale feature decomposition module, a spectral feature extraction module, a topological feature extraction module, and a feature fusion module.
[0043] The multi-scale feature decomposition module: through wavelet transform, the sharpened image is decomposed into high-frequency detail features reflecting edges and small topographic features, and low-frequency approximate features reflecting macroscopic topographic trends. The high-frequency detail features help to identify specific signs and subtle topographic abnormalities of soil erosion, and the low-frequency approximate features provide a basis for analyzing the macro environment and potential risks of soil erosion.
[0044] The spectral feature extraction module: according to the sharpened image, the normalized difference vegetation index (NDVI) and the soil brightness index (SBI) are calculated to generate a spectral feature vector, which represents vegetation coverage and soil characteristics. The higher the NDVI index, the weaker the surface fixation ability, and the higher the risk of soil erosion. The higher the SBI index, the looser or more exposed the soil, and the more easily it is eroded by water flow.
[0045] The topological feature extraction module: a scale space is constructed for the low-frequency approximate features, and the persistent pairs are extracted through persistent homology analysis to generate a topological feature vector, which can quantitatively describe and characterize the connectivity, hole features and spatial topological relationships of the topographic structure. Furthermore, the structural stability of the low-frequency approximate features reflecting the macroscopic topographic trend is further mined, which makes up for the deficiency of traditional geometric analysis in describing the "connectivity-disruption" of the topography.
[0046] The feature fusion module: the high-frequency detail features, low-frequency approximate features, spectral feature vectors, and topological feature vectors are spliced and fused to generate a fusion feature map containing spatial, spectral, and structural information.
[0047] The risk identification module first performs preliminary screening: based on the fused feature map, the soil and water loss probability is calculated by a random forest algorithm, and high-risk areas are marked.
[0048] Further, the high-risk areas are finely identified. First, the soil loss amount is calculated for the high-risk areas by a soil erosion physical model. Then, the features in the fused feature map, the soil loss amount, and external information received by the interface are taken as nodes of the directed edge, and a directed edge probability model is constructed to realize the probability linkage of the features and the risks.
[0049] This system combines spatial scale features (high frequency / low frequency), spectral features (NDVI / SBI), and topological features (terrain structure) to comprehensively depict the signs of soil and water loss from a micro to a macro perspective, thereby improving the richness of feature expression. Machine learning is used for rapid screening: a random forest algorithm is used to realize efficient preliminary screening of a large area, thereby reducing the cost of manual intervention. A physical model is combined for accurate calculation: a directed edge network is combined with a physical model to quantitatively calculate the high-risk areas, thereby avoiding "one-size-fits-all" evaluation and achieving a balance between efficiency and accuracy.
[0050] Moreover, the directed edge network updates the posterior probability in real time based on data, can dynamically respond to environmental changes, realizes dynamic calibration of risks, and achieves dynamic adaptability and intelligence.
[0051] A power transmission line risk assessment method based on image recognition is as follows:
[0052] In the data acquisition stage,
[0053] Multi-source device cooperation: unmanned aerial vehicles, fixed cameras, and manual shooting are combined to realize the combination of high-resolution images, long-term monitoring, and supplementary shooting of key areas, thereby covering different monitoring scenarios. The time, geographic location, and shooting parameters (such as the flight height of the unmanned aerial vehicle) of the images are recorded to provide a benchmark for subsequent registration and analysis.
[0054] In the data preprocessing stage,
[0055] Temporal and spatial registration: including time alignment and geometric correction;
[0056] Time alignment: based on timestamps, multi-source images are matched to realize the spatio-temporal matching of the collected multi-source information.
[0057] Geometric correction: eliminate mismatched points, and use a polynomial transformation model to unify the images to the same coordinate system.
[0058] Noise suppression: a 3x3 window median filter is used. The gray values of each pixel point in the field are arranged in order from small to large, and the middle gray value in the order is used to replace the gray value of the middle pixel point in the field. Finally, all the pixel points in the image are processed in the above manner to obtain a median filter image.
[0059] Feature enhancement: edge sharpening of the image by Laplacian operator to improve the clarity of the feature boundaries.
[0060] Feature extraction stage,
[0061] Multi-scale decomposition: decompose the features by wavelet transform to get high-frequency features reflecting surface details and low-frequency features reflecting terrain trends.
[0062] Spectral feature calculation:
[0063] NDVI: formula [NIR-R] / [NIR+R], where NIR is the near-infrared band and R is the red band. This formula infers the vegetation coverage in the image by the proportion of color light in the image.
[0064] SBI: formula [R+G+B] / 3, where R is the red channel pixel value, G is the green channel pixel value, and B is the blue channel pixel value. This formula calculates the average value of three channel pixel values to represent the brightness characteristics of the soil surface.
[0065] Topology feature extraction:
[0066] Construct a scale space for low-frequency approximation features, extract persistent pairs and vectorize by persistent homology analysis to represent the structural features of terrain connectivity and holes.
[0067] Feature fusion: splice high-frequency details, low-frequency trends, spectral vectors, and topological vectors into a fusion feature map to realize cross-modal fusion of "spatial-spectral-structure" information.
[0068] Risk identification stage,
[0069] Screening model: build a random forest algorithm model, input features include spectral features, terrain features, topological features, wavelet decomposition features, meteorological features, and time features. For each decision tree node, randomly select different features, and use Gini impurity as the evaluation index of feature splitting quality.
[0070] Calculate the probability of soil erosion for all images according to the random forest algorithm model, and mark the areas with threshold > 70% as high-risk areas.
[0071] Subsequent detailed risk identification,
[0072] First, the directed edge is constructed: define the node association. In specific use, "rainfall intensity→soil loss" and "vegetation coverage→soil erosion probability" are used as a directed edge, in which the nodes are associated with each other and affect each other through a certain direction. All directed edges are directed to the risk level. The nodes are also divided into the basic layer, the intermediate layer and the output layer. The basic layer includes spectral features, topological features, meteorological data and other basic information. The intermediate layer includes soil loss, vegetation coverage and other derived indicators. The output layer is the risk level, and the number of output layer nodes is the risk level classification. The direction of the directed edge is along the direction of the basic layer-intermediate layer-output layer.
[0073] At the same time, the construction logic of the directed edge follows the determination of nodes and directions based on the causal relationship and data correlation.
[0074] Conditional probability table: first discretize the continuous quantity, including soil erosion probability, soil loss, NDVI and slope output by the random forest algorithm model, etc. After discretization, the state quantity of the node is reduced, which is convenient for subsequent probability statistical processing, and then the conditional probability between nodes is trained based on historical data through maximum likelihood estimation using historical labeled data.
[0075] Inference calculation: input real-time data, update the posterior probability of each node according to the inference network composed of all directed edges, and calculate the probability value of the output node through joint probability distribution.
[0076] Risk level as the final direction of the directed edge, through numerical classification, as the final output.
[0077] The preprocessing process of this method is through registration→denoising→sharpening to form a standardized link, which can adapt to different sensor data and improve the universality of the method.
[0078] Median filtering and geometric correction effectively suppress noise and distortion, ensuring the reliability of subsequent feature extraction and improving the standardization and robustness of data processing
[0079] And through spatial features (wavelet decomposition) to capture the changes of the surface form, spectral features to reveal the physical properties of vegetation and soil, and topological features to depict the stability of the terrain structure, the three types of features complement each other, realizing the complementarity of multi-dimensional features and avoiding misjudgment of a single feature.
[0080] In actual use, NDVI is low (sparse vegetation) + high-frequency features exist abnormal edge (may be erosion ditch) + topological features show poor terrain connectivity (easy to occur landslide), and the combination of the three can significantly improve the recognition accuracy of high-risk areas.
[0081] Through model combination, the efficiency of random forest is used: high-dimensional features (fusion feature maps contain hundreds of features) are quickly processed through parallel decision trees to achieve a large-scale "rough screening", which is more than 50% more efficient than traditional physical models; combined with the explainability of directed edge network: through probability reasoning, the influence path of each factor on soil erosion is clear, such as rainfall → soil moisture → erosion, the output result is traceable, and it is convenient for operation and maintenance personnel to understand the cause of the risk.
[0082] And support real-time data input, when heavy rain warning, can update the rainfall node state in time, trigger the risk level dynamic adjustment, realize the "monitoring-analysis-warning" closed loop.
[0083] The above description of disclosed embodiments enables those skilled in the art to carry out or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. An image recognition based power line risk assessment system, characterized in that, The system comprises a multi-source data acquisition module, a data preprocessing module, a feature extraction module, and a risk identification module. The multi-source data acquisition module collects external image information, account data, historical maintenance records, and real-time meteorological factors, and transmits the image data to the data preprocessing module for preprocessing, and transmits the preprocessed image data to the feature extraction module to obtain a fusion feature map of the image, and finally the risk identification module judges the soil erosion risk level of the image through the fusion feature map.
2. The image recognition based transmission line risk assessment system of claim 1, wherein, The data preprocessing module comprises a space-time registration module, a noise suppression module, and a feature enhancement module. The space-time registration module receives image information from the multi-source data acquisition module, and transmits the registered image to the noise suppression module, and the de-noised image is sharpened by the feature enhancement module and outputted as a sharpened image.
3. The image recognition based transmission line risk assessment system of claim 1, wherein, The feature extraction module comprises a multi-scale feature decomposition module, a spectral feature extraction module, a topological feature extraction module, and a feature fusion module. The multi-scale feature decomposition module receives the sharpened image for feature decomposition to obtain high-frequency detail features and low-frequency approximation features; the spectral feature extraction module receives the sharpened image and outputs a spectral feature vector; the topological feature extraction module receives the low-frequency approximation feature map and outputs a topological feature vector; and the feature fusion module splices the high-frequency detail features, the low-frequency approximation features, the spectral feature vector, and the topological feature vector to obtain a fusion feature map.
4. A method for image recognition-based risk assessment of a power transmission line, the method being performed with the aid of a system according to any one of claims 1-3, characterized in that, The method comprises the following steps: In the data acquisition stage, image information is obtained by manual shooting, fixed camera shooting, and unmanned aerial vehicle shooting, and account data, historical maintenance records, and real-time meteorological factors are received through an interface; In the data preprocessing stage, the image information and non-image information obtained by the data acquisition module are first aligned according to time stamps, and a coordinate system is unified through geometric correction; then the registered image is windowed and de-noised, and the de-noised image is processed by the feature enhancement module through a second-order differential sharpening algorithm to obtain a sharpened image; In the feature extraction stage, the sharpened image is decomposed into high-frequency detail features reflecting edges and small landforms and low-frequency approximation features reflecting macroscopic terrain trends through wavelet transform, and normalized vegetation index and soil brightness index are calculated to obtain NDVI values and SBI values; further, a scale space is constructed for the low-frequency approximation features, persistent pairs are extracted, and are vectorized into a topological feature vector; finally, multi-modal features are spliced to obtain a fusion feature map; In the risk identification stage, the fusion feature map is first screened, the soil erosion occurrence probability is calculated through a random forest algorithm, and the high-risk area is marked; the soil loss amount is further calculated for the high-risk area, and the risk level is determined through the fusion feature data.
5. The image recognition-based power line risk assessment method according to claim 4, characterized in that, In the risk identification stage, the specific process of determining the risk level through the fusion feature data is as follows: Construction of directed edges, correlation of input data, determination of directed edge nodes and directions; the directed edges all point to the risk degree nodes finally Construction of conditional probability table, first discretization of continuous quantities in the input data, and then definition of conditional probability between nodes through historical data; Inference calculation, according to real-time data, updates the posterior probability of each node, and calculates the posterior probability of the output node according to the probability of the intermediate node.