Method and System for Identifying Abnormal Areas of Transmission Lines by Fusing Infrared Thermal Image Features

Abstract

The present invention relates to the technical field of power system monitoring, and discloses a method for identifying abnormal regions of transmission lines by infrared thermal image feature fusion, including the following steps: collecting infrared thermal image data of the transmission line and preprocessing the data; performing multi-scale decomposition on the preprocessed infrared thermal image through wavelet transform to extract the low-frequency part and the high-frequency part. The present invention also provides a system for identifying abnormal regions of transmission lines by infrared thermal image feature fusion, including: a data acquisition module for collecting infrared thermal image data of the transmission line; a preprocessing module for preprocessing the collected infrared thermal image data. By combining multi-scale wavelet transform, nonlinear dynamics modeling, graph neural network optimization, and reinforcement learning, the present invention improves the accuracy of abnormal region identification, enhances the robustness and adaptability of the system in complex environments, and ensures efficient and accurate identification of abnormal regions under dynamic changes and noise interference.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring, and specifically to a method and system for identifying abnormal areas of transmission lines by fusing infrared thermal image features. Background Art

[0002] With the expansion of the scale of the power system and the increase in power demand, the health monitoring of transmission lines has become increasingly important. Equipment failures or temperature anomalies in transmission lines may trigger serious power accidents, affecting social production and residents' lives. Therefore, it is crucial to promptly detect and handle the abnormal states of transmission lines to ensure the stability and security of the power system. As a non-contact temperature detection method, infrared thermal imaging technology can monitor the temperature changes on the surface of equipment in real time through an infrared camera, and can quickly identify potential abnormal areas such as overheating and faults.

[0003] In the prior art, the processing of infrared thermal images mainly relies on traditional image analysis techniques, such as threshold-based anomaly detection methods and edge detection algorithms. These methods can determine whether an equipment is abnormal through a simple temperature threshold and perform edge detection on the image to identify potential fault areas. In some simple monitoring scenarios, traditional methods can effectively identify abnormal temperature areas and, to a certain extent, help maintenance personnel quickly locate faults. These technologies have realized the automation of temperature monitoring, reduced the workload of manual inspections, and improved the efficiency of preliminary fault diagnosis.

[0004] Although the prior art has achieved certain success in some application scenarios, its disadvantages are also obvious; firstly, the existing threshold-based detection methods lack flexibility and cannot adapt to temperature changes in different environments. Due to the continuous changes in equipment and environmental conditions, the accuracy of these methods is greatly reduced when dealing with complex or variable environments; secondly, the prior art mostly relies on static features, such as temperature thresholds or the overall brightness distribution of images, and ignores the dynamic and spatio-temporal evolution characteristics of temperature changes. Problems such as overheating and faults in equipment are usually a continuous process of change, and traditional methods fail to consider these spatio-temporal characteristics, resulting in limited recognition accuracy; thirdly, the existing methods have a weak response to subtle changes or local faults in images and are prone to missing relatively small abnormal areas, affecting the overall performance of the system; finally, although some modern image processing technologies (such as deep learning) have begun to be applied to infrared thermal image analysis, most of the existing methods rely on a large amount of labeled data for training and have poor adaptability to environmental changes, and there is still a large room for optimization. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a method and system for identifying abnormal areas of transmission lines by fusing infrared thermal image features, which solves the problems of low accuracy in identifying abnormal areas, insufficient capture of dynamic changes, and lack of adaptive optimization in the existing technology in complex environments.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for identifying abnormal areas of transmission lines by fusing infrared thermal image features, comprising the following steps:

[0007] Collect infrared thermal image data of the transmission line and preprocess the data;

[0008] Perform multi-scale decomposition on the preprocessed infrared thermal image through wavelet transform, and extract the low-frequency part and the high-frequency part;

[0009] Embed the extracted multi-scale features through a manifold learning method to obtain a feature set in the low-dimensional space;

[0010] Model the spatio-temporal evolution of the abnormal area based on a non-linear dynamics model and fuse it with the embedded feature set;

[0011] Use a graph neural network to optimize the fused features, update the relationships between nodes, and output the optimized features;

[0012] Utilize the optimized features to identify the abnormal area and obtain the location of the abnormal area;

[0013] Optimize the abnormal area recognition strategy through reinforcement learning and a global optimization algorithm to improve the accuracy of the recognition result.

[0014] Preferably, the preprocessing includes:

[0015] Remove noise from the collected infrared thermal image data;

[0016] Perform contrast enhancement processing on the infrared thermal image;

[0017] Perform histogram equalization operation on the processed image.

[0018] Preferably, the wavelet transform adopts discrete wavelet transform. The low-frequency part includes the overall temperature distribution characteristics, and the high-frequency part includes the detail characteristics. The extraction process of the low-frequency part and the high-frequency part includes:

[0019] Perform wavelet transform on the image data to obtain low-frequency and high-frequency features at different scales;

[0020] Denoise the low-frequency and high-frequency parts and retain the parts with significant features.

[0021] Preferably, the high-dimensional data embedding is performed using the principal component analysis method or the t-SNE method, and the feature set after data embedding includes:

[0022] Perform dimensionality reduction on the high-dimensional data to extract the most representative features;

[0023] Use the dimensionality-reduced feature set as the input for subsequent processing.

[0024] Preferably, the non-linear dynamics modeling uses the Lorenz system to describe the dynamic evolution process of the abnormal region, and the dynamic evolution process includes:

[0025] Based on the non-linear equation of the Lorenz system, simulate the dynamic evolution process of the abnormal region in the infrared thermal image over time. By using the three-dimensional state variables of the Lorenz equation, describe the dynamic processes of temperature change, hot spot expansion and contraction in the abnormal region;

[0026] In the Lorenz system, by selecting appropriate parameters and adjusting according to the actual infrared thermal image data, make the model better match the actual temperature change pattern;

[0027] Fuse the dynamic evolution process obtained by Lorenz system modeling and the multi-scale features extracted from the infrared thermal image through a weighted fusion method.

[0028] Preferably, the optimization of the graph neural network includes:

[0029] Take each image feature node as a node of the graph, and represent the similarity between nodes through the edges of the graph;

[0030] Normalize the adjacency matrix of the graph, use the graph neural network to propagate the features, and update the node features;

[0031] Use graph convolution operations to integrate the node features to obtain an optimized feature set as the final output.

[0032] Preferably, the abnormal region recognition uses the support vector machine or decision tree method, and the recognition results include:

[0033] Build a classification model based on the training data for automatic recognition of the abnormal region;

[0034] Output the position coordinates and temperature information of the recognized abnormal region.

[0035] Preferably, the reinforcement learning algorithm includes:

[0036] Define the state space, action space and reward function;

[0037] Use Q-learning or deep Q-network to optimize the abnormal region recognition strategy;

[0038] The strategy is adjusted according to the feedback reward function to maximize the recognition accuracy.

[0039] Preferably, the global optimization uses a genetic algorithm or a simulated annealing algorithm to globally optimize the identification strategy, and the optimization process includes:

[0040] Searching for parameters of the identification strategy using genetic algorithms or simulated annealing algorithms;

[0041] The stability and accuracy of the strategy are optimized through global search.

[0042] The present invention also provides a transmission line abnormal area recognition system using infrared thermal image feature fusion, comprising:

[0043] Data acquisition module, used to collect infrared thermal imaging data of power transmission lines;

[0044] A preprocessing module, used for preprocessing the collected infrared thermal imaging data;

[0045] Feature extraction module, used to perform wavelet transform, multi-scale analysis, data embedding and nonlinear dynamic modeling on the pre-processed infrared thermal imaging data to extract multi-dimensional features;

[0046] Graph neural network optimization module, used to optimize the graph structure of the extracted features;

[0047] An abnormal area recognition module is used to locate and identify abnormal areas based on optimized features;

[0048] Reinforcement learning and global optimization modules are used to optimize abnormal area identification strategies and improve identification accuracy.

[0049] The present invention provides a method and system for identifying abnormal areas of power transmission lines by integrating infrared thermal imaging features. It has the following beneficial effects:

[0050] 1. The present invention extracts multi-scale features of images through wavelet transform, combines the Lorenz system for nonlinear dynamic modeling, and accurately captures static and dynamic information in the image. Compared with traditional methods that only focus on static features of images, the present invention can dynamically analyze the temperature changes and expansion process of abnormal areas, significantly improving the accuracy and comprehensiveness of abnormal area identification.

[0051] 2. The present invention optimizes the relationship between feature nodes through graph neural networks, so that each feature in the image can better integrate contextual information. Different from traditional feature processing methods, graph neural networks can effectively mine the spatial and temporal dependencies between image features, greatly improving the accuracy of locating abnormal areas, especially in complex environments.

[0052] 3. The present invention optimizes the recognition strategy through reinforcement learning and combines global optimization methods such as genetic algorithms, enabling the system to adapt to different environmental conditions. In traditional methods, the strategy is fixed and vulnerable to environmental changes. In contrast, the present invention improves the robustness and accuracy of the system in complex environments through continuous learning and adjustment. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is the flowchart of the method of the present invention;

[0054] Figure 2 is the system structure diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the specification of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0056] Please refer to the attached Figure 1 , the embodiments of the present invention provide a method for identifying abnormal areas of transmission lines by fusing infrared thermal image features, including the following steps:

[0057] S1. Collect and preprocess data: Collect infrared thermal image data of the transmission line and preprocess the data;

[0058] This step lays the foundation for subsequent multi-scale feature extraction, dynamic modeling, and abnormal area recognition. Data collection and preprocessing directly affect the performance of subsequent modules, ensuring data accuracy and robustness during subsequent step processing. To overcome possible influences such as environmental noise and image distortion, a series of effective image processing techniques are adopted in this step to improve data quality.

[0059] In this embodiment, temperature distribution data of the transmission line is collected by using a high-precision infrared thermal imager. Infrared thermal imagers usually utilize the principle of infrared radiation and can accurately measure and record the temperature information on the surface of power equipment. The collected infrared thermal image is usually a two-dimensional image, and each pixel point corresponds to the temperature of different areas on the surface of the equipment.

[0060] After image collection, in order to eliminate noise and interference during the collection process, the image must be preprocessed. Generally, image preprocessing includes operations such as noise removal, contrast enhancement, and histogram equalization. These operations can significantly improve image quality, eliminate the influence of environmental factors on data, and thus improve the accuracy of subsequent analysis processes.

[0061] During the process of infrared thermal image data acquisition, due to environmental changes or equipment limitations, the image may contain certain noise. This noise mainly comes from the electronic noise of the sensor, interference signals during image transmission, and external environmental factors (such as temperature fluctuations, weather changes, etc.). To solve this problem, the present invention uses Gaussian filtering or median filtering to remove the noise.

[0062] Gaussian filtering: Generally, Gaussian filtering is used in the process of image denoising and can smooth the noise in the image. The formula of the Gaussian filter is:

[0063] ;

[0064] where, is the value of the Gaussian filter, is the standard deviation, representing the width of the filter and controlling the degree of denoising; and are the pixel coordinates.

[0065] Median filtering: As an option, median filtering can effectively remove salt-and-pepper noise and retain the image edge information. The basic principle of median filtering is to replace the target pixel value by taking the median of the neighboring pixels.

[0066] The temperature change of the infrared thermal image is often relatively subtle. Especially in a complex background environment, the abnormal area may be difficult to identify. Therefore, contrast enhancement can highlight the detailed parts in the image, making the abnormal area more prominent. In this embodiment, the histogram equalization technique is used to enhance the contrast of the image.

[0067] Histogram equalization is a commonly used image enhancement technique. Its purpose is to make the gray level range of the image more uniform by adjusting the gray level distribution of the image. Specifically, this method improves the contrast of the image by equalizing the cumulative distribution function (CDF) of the image, thereby enhancing the local details of the image. Its formula is as follows:

[0068] ;

[0069] where, is the cumulative distribution function of the image gray level , is the probability density function of the gray level in the image; : represents a gray level or brightness value in the image; : represents the currently accumulated gray value.

[0070] While performing contrast enhancement, histogram equalization redistributes the grayscale of the image so that the brightness range of the image is more evenly distributed. Through this processing, the temperature differences in the image will be more obvious and abnormal areas will be easier to detect.

[0071] In one possible implementation, multiple enhancement techniques may be combined to enhance image details. For example, after noise removal, an adaptive contrast enhancement method is used to adaptively adjust contrast in a local area to ensure that abnormal areas are clearly represented under various environmental conditions.

[0072] In this embodiment, the infrared thermal image after the above preprocessing will enter the wavelet transform module. The image after denoising and enhancement has a higher signal-to-noise ratio and clearer temperature distribution information, which can provide high-quality input for subsequent multi-scale wavelet transform and feature extraction. By further decomposing the image, features at different scales are extracted, providing more accurate basic data for subsequent feature fusion and abnormal area identification.

[0073] Through this series of preprocessing, the system can effectively reduce the impact of environmental interference and equipment errors, while improving the detection effect of abnormal areas, thereby ensuring the accuracy and robustness of the entire abnormal area identification method.

[0074] Through accurate acquisition and multi-level preprocessing of infrared thermal imaging data, noise is removed, contrast is enhanced, and grayscale distribution of the image is balanced. This step ensures the efficiency and reliability of subsequent feature extraction and abnormal area identification, and provides accurate input data for the entire system.

[0075] S2. Multi-scale wavelet transform feature extraction: The pre-processed infrared thermal image is decomposed into multiple scales through wavelet transform to extract the low-frequency part and the high-frequency part;

[0076] Multi-scale wavelet transform can effectively extract information from images at different frequency scales, ensuring that the system can perform effective detection regardless of the size or location of the abnormal area.

[0077] In this embodiment, after the preprocessing steps (such as denoising, contrast enhancement, and histogram equalization), the infrared thermal image already has a higher signal-to-noise ratio and clearer temperature distribution information. The application of wavelet transform at this stage can further explore the deeper local and global features in the image. Through wavelet transform, the low-frequency and high-frequency parts of the image are separated to facilitate more effective processing of features of different scales.

[0078] In general, as a classic signal processing tool, wavelet transform is widely used in image analysis, especially in scenarios where information at different scales needs to be captured. Its main advantage lies in being able to provide localized information in both space and frequency, which is very helpful for identifying abnormal regions in infrared thermal images of transmission lines, especially the edges and local faults.

[0079] Specifically, in this embodiment, the image is first decomposed by discrete wavelet transform (DWT) to extract the low-frequency part and the high-frequency part. The low-frequency part contains the overall temperature change of the image, while the high-frequency part contains the local details of the image and the minute fluctuations of the temperature change. The high-frequency part usually corresponds to the hot spots, edges or mutation regions in the image, which is crucial for locating and identifying abnormal regions.

[0080] In a possible implementation, wavelet transform decomposes the image into multiple scales to capture features at different scales. The low-frequency part represents the rough structure of the image, usually the region with relatively uniform temperature, while the high-frequency part reflects the region with drastic temperature change, usually corresponding to the region where problems or faults occur in electrical equipment.

[0081] The specific implementation process includes:

[0082] The low-frequency part of wavelet transform generally corresponds to the smooth region of the image. In infrared thermal images, the low-frequency part usually reflects the overall temperature distribution. The temperature change in these regions is relatively slow, usually indicating that there is no serious fault in the electrical equipment. The extraction of the low-frequency part can help the system better understand the overall temperature trend and provide a reference for subsequent abnormal region detection.

[0083] The high-frequency part reflects the local details in the image, especially the mutation regions, such as equipment faults, hot spots or current overloads. The hot spot region usually shows a phenomenon of sudden temperature rise. By extracting the high-frequency part through wavelet transform, the characteristics of these abnormal regions can be effectively highlighted. The high-frequency part is usually used to detect regions with drastic temperature change and local abnormalities.

[0084] To further improve the accuracy of abnormal region recognition, the multi-scale wavelet transform technology is adopted, which can extract features simultaneously at multiple scales. The multi-scale transform can ensure that no matter the size of the abnormal region, the system can effectively capture the corresponding details. Wavelet transform is not limited to a single layer of decomposition, but through multiple scale conversions, gradually extracts detailed information at different levels.

[0085] In another embodiment, the image is gradually decomposed into smaller and smaller parts by means of multi-layer wavelet transform to facilitate capturing detailed features. After each decomposition, the system will retain the corresponding features according to different scales of temperature change for subsequent processing to avoid losing key details.

[0086] Through this process, the low-frequency part and the high-frequency part represent different physical phenomena respectively. The low-frequency part mainly helps to identify the normal state of the device, while the high-frequency part helps to identify the areas where faults may exist. For infrared thermal imaging data, the low-frequency part reflects the average trend of the temperature distribution and can reveal the overall situation; while the high-frequency part reflects the fault signs such as temperature anomalies and local overheating.

[0087] In this embodiment, the extracted multi-scale features will be used as inputs and enter the subsequent feature fusion and non-linear dynamics modeling steps. These features provide a solid foundation for the detection of abnormal areas. Through subsequent feature fusion and modeling, the recognition accuracy of abnormal areas can be further improved. The multi-scale wavelet transform provides rich feature information for the system, ensuring that the system can effectively handle various temperature anomalies in complex environments.

[0088] Through multi-scale wavelet transform, the present invention can extract key features in infrared thermal imaging images from different scales, including the overall temperature change in the low-frequency part and the local details in the high-frequency part. This step effectively enhances the detection ability of the system for abnormal areas in complex environments, improves the recognition accuracy, and ensures the ability to capture details at different scales.

[0089] S3. High-dimensional feature reduction: Embed the extracted multi-scale features into high-dimensional data through a manifold learning method to obtain a set of features in the low-dimensional space;

[0090] Feature reduction can not only reduce the processing difficulty of data, but also remove redundant information, retain the most representative features, and further improve the efficiency and accuracy of abnormal area recognition.

[0091] In this embodiment, when performing high-dimensional feature reduction, we adopt manifold learning techniques, especially principal component analysis (PCA) and t-SNE methods to reduce the dimension of the data. These methods can effectively map high-dimensional data into a low-dimensional space while retaining as much important information in the original data as possible.

[0092] Generally, high-dimensional data usually contains a lot of redundant or noise information. The purpose of feature reduction is to map the data from a high-dimensional space to a lower-dimensional space, making the important feature information more concentrated and facilitating subsequent processing. Through feature reduction, the complexity of the data can be significantly reduced, the computational efficiency can be improved, and it helps to avoid the overfitting problem.

[0093] Specifically, in this embodiment, we first use the principal component analysis (PCA) method to perform preliminary dimensionality reduction on the features. PCA is a linear dimensionality reduction technique that reduces the dimensionality of data by projecting the data onto new coordinate axes and selecting the most representative principal components. The basic principle of PCA is to perform eigenvalue decomposition on the data covariance matrix and select the principal components with larger eigenvalues as the new features, thereby retaining most of the information in the data.

[0094] The calculation process of PCA is as follows:

[0095] ;

[0096] Where: is the original high-dimensional feature matrix; is the projection matrix composed of eigenvectors; is the feature matrix after dimensionality reduction.

[0097] In this process, PCA maximally retains the part with larger variance in the data through the linear combination of eigenvectors, thereby retaining the most important information. The explanatory power of each principal component can be measured by the magnitude of its corresponding eigenvalue, and the principal component corresponding to a larger eigenvalue represents more important changes in the data.

[0098] As an option, if it is necessary to further capture the complex non-linear relationships in the data, the t-SNE (t-Distributed Stochastic Neighbor Embedding) method can be adopted, which is particularly suitable for dealing with complex non-linear relationships in high-dimensional data. t-SNE embeds the data into a low-dimensional space by calculating the similarity between high-dimensional data points and using probability distributions, so that similar data points in the high-dimensional space are also as close as possible in the low-dimensional space.

[0099] The dimensionality reduction process of t-SNE is achieved in the following way:

[0100] Calculate the conditional probability of each pair of data points to measure the similarity between them;

[0101] Maintain the local structure in the high-dimensional data by minimizing the KL divergence between similar points in the low-dimensional space.

[0102] Specifically, in this embodiment, PCA is mainly used for preliminary dimensionality reduction, reducing the high-dimensional features to a lower dimension and retaining the most representative principal components. While t-SNE can more effectively maintain the local structure when further reducing the dimension, especially in complex non-linear data. The data after dimensionality reduction reduces noise interference by retaining the most important feature information, which helps subsequent modeling and identification of abnormal regions.

[0103] In a possible implementation, the dimension-reduced feature data not only reduces the dimension and computational cost but also retains the essential features of the data. These dimension-reduced data will be used as the input for subsequent non-linear dynamics modeling and graph neural network optimization to further improve the accuracy and efficiency of abnormal region recognition.

[0104] Through dimension reduction, the originally high-dimensional and complex data is compressed into simpler and more structured low-dimensional data, removing redundant information. This process significantly improves the efficiency of subsequent algorithms (such as non-linear modeling and graph neural network optimization), and also avoids the curse of dimensionality problem that may occur in high-dimensional spaces, enabling the entire system to perform efficient processing at a lower computational cost.

[0105] The dimension-reduced feature set will enter the non-linear dynamics modeling step. In this step, the dimension-reduced data not only reduces the computational complexity but also effectively retains the important information in the image, enabling the subsequent non-linear modeling based on the Lorenz system to more accurately and efficiently capture the dynamic evolution of the abnormal region.

[0106] Generally speaking, through high-dimensional feature dimension reduction, the present invention can significantly improve the data processing efficiency while maintaining the key information of the image, providing high-quality input data for subsequent model construction and abnormal region detection.

[0107] By reducing the dimension of the high-dimensional feature data, the data dimension and computational burden are reduced, while the most important information features are retained.

[0108] S4. Non-linear dynamics modeling and feature fusion: Model the spatio-temporal evolution of the abnormal region based on the non-linear dynamics model and fuse it with the embedded feature set;

[0109] Through non-linear dynamics modeling and feature fusion, the ability to model the spatio-temporal evolution of the abnormal region is further enhanced, and the features from different modules are effectively combined to improve the accuracy of subsequent abnormal region recognition.

[0110] In this embodiment, we use the Lorenz system as the non-linear dynamics model to simulate the spatio-temporal evolution process of the abnormal region in the infrared thermal image. Since the temperature change of the abnormal region is often affected by various factors, the Lorenz system can effectively model these changes and fuse with the dimension-reduced feature set on this basis. In this way, the system can not only identify the static features in the image but also capture the dynamic features of the abnormal region changing over time.

[0111] Generally, when dealing with infrared thermal images, the changes in abnormal regions are usually non-linear and dynamic, and may be affected by multiple factors such as temperature, humidity, and current load. These factors may cause a sharp increase or unstable change in local temperature, while traditional static feature analysis methods often fail to effectively capture these changes. By introducing the Lorenz system, we can better simulate these complex non-linear processes, especially in the context of multi-dimensional data and dynamic changes, with stronger adaptability and accuracy.

[0112] Specifically, the Lorenz system consists of three equations, which are used to describe a three-dimensional non-linear system. Its basic form is as follows:

[0113] ;

[0114] ;

[0115] ;

[0116] Where: , , respectively represent the three state variables of the Lorenz system. In the present invention, represents the temperature change, represents the speed of hot spot expansion, represents other relevant temperature factors or equipment states; , , are the parameters of the Lorenz system, usually set as the initial conditions of the system. By adjusting these parameters, different temperature change patterns and equipment states can be simulated.

[0117] As an option, in this system, we use a numerical solution method to solve the above differential equations, and train and optimize the system based on historical data. By continuously adjusting , and parameters, it can better adapt to the changes in abnormal regions under different environments, thereby improving the fitting accuracy of the model.

[0118] In a possible implementation, by fusing with the reduced-dimensional feature set, the Lorenz model can provide time series data for each abnormal region, and these data reflect the dynamic characteristics of the region changing over time. These dynamic characteristics combined with static image features (such as temperature distribution) can provide more comprehensive information for subsequent abnormal region recognition.

[0119] Fusing Static Features and Dynamic Features: By means of weighted fusion, the static features extracted by wavelet transform (such as low-frequency temperature distribution) are combined with the dynamic features obtained from Lorenz system modeling. The weights are set based on the contribution of each feature to the identification of abnormal regions. Dynamic features usually account for a larger proportion in abnormal regions with obvious temporal changes, while static features are more used to describe the overall temperature trend.

[0120] The Fused Feature Set: With the fused feature set, the system can simultaneously consider the temperature changes and temporal evolution process in the image. This fusion not only improves the accuracy of identification but also enhances the ability to identify abnormal regions in complex scenarios.

[0121] Specifically, the fused feature set can accurately describe two important dimensions of abnormal regions in the image: spatial features (temperature distribution) and temporal features (dynamic changes of abnormal regions). This combination greatly enhances the adaptability of the system in various environments and can still effectively identify potential fault regions when the temperature change is not obvious or the device state changes slowly.

[0122] In another implementation, to further improve the robustness of the system, more dynamic modeling methods can be introduced into the fusion process, such as combining time series analysis (such as ARIMA model) to further optimize the prediction of abnormal regions. By combining multiple models, more accurate results can be obtained under different conditions.

[0123] In this embodiment, nonlinear dynamics modeling and feature fusion provide highly accurate dynamic features, which provide key inputs for the subsequent optimization of the graph neural network. Through this process, the system can not only extract important features from static images but also predict the evolution trend of abnormal regions based on historical data and dynamic changes.

[0124] By using the Lorenz system for nonlinear dynamics modeling, the dynamic changes of abnormal regions can be accurately captured and these dynamic features can be effectively fused with static features. This step significantly improves the system's ability to identify abnormal regions in complex environments and ensures high-precision and high-robustness detection of abnormal regions under different working conditions. Through feature fusion, the system can simultaneously consider the temperature distribution and the evolution process of temperature changes, providing comprehensive and accurate feature data for subsequent identification of abnormal regions.

[0125] S5. Optimize Features with Graph Neural Network: Use the graph neural network to optimize the fused features, update the relationships between nodes, and output the optimized features;

[0126] This step utilizes the graph structure information of the graph neural network to optimize the relationships between each feature node, further improving the quality and distinguishability of feature representation.

[0127] In this embodiment, the main task of optimizing the graph neural network is to update the features of each node in the graph through graph convolution operations, so that the features can better reflect the spatial structure and local dependence relationship of the abnormal region. Specifically, the graph neural network regards each feature in the infrared thermal image as a node, and the relationship between different feature nodes is represented by the edges of the graph. Through graph convolution operations, information between nodes is transmitted, and node features are continuously updated, enabling the network to optimize features on the graph structure.

[0128] Generally, the graph neural network can effectively process graph-structured data and optimize the mutual relationship between nodes through local information transmission. For infrared thermal images, the advantage of the graph neural network is that it can capture the spatial and temporal dependencies between different temperature changes, especially the local features of the abnormal regions in the image. The graph neural network propagates information between nodes through convolution operations, making the features of each node depend not only on its own attributes but also on the attributes of neighboring nodes.

[0129] Specifically, the basic structure of the graph neural network includes an adjacency matrix of the graph and a node feature matrix. The adjacency matrix represents the connection relationship between nodes in the graph, and the node feature matrix represents the features of each node. The graph convolution operation is based on the adjacency matrix and node features for information propagation and feature update.

[0130] The general form of the graph convolution operation is:

[0131] ;

[0132] where: is the node feature matrix of the th layer; is the normalized adjacency matrix, representing the similarity relationship between nodes; is the weight matrix of the th layer, used to map the feature dimensions; is the activation function, and commonly used activation functions include ReLU or sigmoid; is the updated node feature matrix of the th layer.

[0133] In a possible implementation, the adjacency matrix is usually normalized to avoid the influence of the degree difference of nodes in the graph on feature propagation. The normalized adjacency matrix can be obtained through the following formula:

[0134] ;

[0135] where, is the degree matrix of the nodes in the graph, It is the original adjacency matrix of the graph. Through this standardization, the propagation process of node features is balanced, avoiding unbalanced influences between nodes.

[0136] As an option, in some cases, the convolutional operation of the graph neural network can also be strengthened by stacking multiple layers to enhance the feature learning ability. Specifically, through multi-layer graph convolutional operations, the network can capture a wider range of relationships between nodes and deep feature dependencies, thereby improving the recognition ability for abnormal regions in complex scenarios.

[0137] Specifically, the main task of the graph neural network to optimize features is to update the information layer by layer, so that the features of each node can more accurately reflect the spatial structure of the abnormal region in the image. For example, in infrared thermal images, temperature changes are usually local, and abnormal regions usually appear as concentrated hot spots. The graph neural network can capture this local structure, enabling the network to optimize features for specific abnormal regions and further improving the localization accuracy of abnormal regions.

[0138] In a possible implementation, to further enhance the graph neural network's learning ability for relationships between nodes, the multi-head attention mechanism (Multi-Head Attention) can be adopted. This mechanism allows the network to learn the relationships between nodes in parallel in multiple subspaces, thereby enhancing the network's expressive ability. The multi-head attention mechanism can help the network better handle complex dependency relationships, especially when dealing with images containing a large number of nodes and complex edge relationships, which can improve the performance of the graph neural network.

[0139] Through the optimization of the graph neural network, the features in the image will be able to be recombined according to the similarity between nodes, enabling each feature node to more accurately reflect the spatial distribution and local features of the abnormal region. This process not only improves the accuracy of feature representation but also enhances the robustness of the system, enabling the system to still maintain a high recognition accuracy in a complex environment.

[0140] The feature optimization of the graph neural network provides accurate and optimized feature inputs for subsequent abnormal region recognition steps. These optimized features can more accurately reflect the abnormal regions in the image, providing high-quality data support for subsequent classification and localization.

[0141] S6. Abnormal region recognition: Use the optimized features to recognize the abnormal region and obtain the location of the abnormal region;

[0142] The identification of abnormal regions uses these features to accurately locate possible abnormal regions in the image, such as hot spots with excessive temperature or equipment failures, etc. This step is the core of the entire system, determining whether the system can efficiently and accurately identify the regions that need attention and providing a basis for subsequent power equipment maintenance.

[0143] In this embodiment, the identification of abnormal regions adopts classical classification algorithms such as Support Vector Machine (SVM) or decision tree. By using these machine learning algorithms, the system can learn the characteristic patterns of abnormal regions based on the optimized feature data and accurately identify the positions and related information of abnormal regions in new images. Through this step, the system can screen out the regions that actually need attention from complex infrared thermal images, avoiding false alarms and missed detections.

[0144] Generally, the identification of abnormal regions relies on the differences between features. There are usually some regions with higher or lower temperatures in the image, and these regions may be related to problems such as overload, local faults, and hot spots of power equipment. Classification algorithms such as Support Vector Machine (SVM) and decision tree can accurately distinguish normal regions and abnormal regions by learning the features of abnormal regions in historical labeled data.

[0145] Specifically, in this embodiment, a classification model is first established through Support Vector Machine (SVM) or decision tree. The input of this model is the feature data extracted and optimized in the previous steps, including the low-frequency and high-frequency features after wavelet transform, the dynamic features in the nonlinear dynamics model, and the features optimized through graph neural network. These features are used to describe information such as the temperature distribution and change trend of different regions in the image.

[0146] The goal of Support Vector Machine (SVM) is to maximize the margin between classifications by finding a hyperplane, thereby dividing the data into two categories. During the training process, SVM solves the following optimization problem:

[0147] ;

[0148] The constraint conditions are:

[0149] ;

[0150] Where: is the normal vector of the hyperplane; is the bias of the hyperplane; is the feature vector of the th sample point; is the label of the th sample; is the slack variable, allowing some sample points to violate the classification boundary to improve the robustness of the model; is a regularization parameter used to control the complexity of the model.

[0151] As an option, in addition, decision tree algorithms can also be used for anomaly region recognition. Decision trees learn the relationship between feature values and target values in the training data to form a tree structure, and finally determine which class a sample belongs to through leaf nodes. During the training process, decision trees select the optimal splitting feature through information gain or Gini index and gradually construct the tree structure. The learning process of decision trees can be expressed by the following formula:

[0152] ;

[0153] where: is the dataset whose information entropy represents the purity of the data; is the dataset split by feature to obtain subsets; and respectively represent the sizes of the dataset and the subset ; is the information entropy of the subset .

[0154] Specifically, the training processes of support vector machines and decision trees require a large amount of labeled data to learn the boundary between normal regions and anomaly regions. Through these machine learning methods, the system can extract the features of anomaly regions from the training data, generate a classification model, and then classify new images in actual applications to accurately identify anomaly regions.

[0155] In a possible implementation, to improve the accuracy of anomaly region recognition, the recognition process can also combine multi-classification algorithms, such as expanding a traditional binary classification model into a multi-classification problem to handle different types of anomaly regions. Through the multi-classification model, the system can not only identify ordinary anomaly regions but also distinguish different types of anomalies (such as overload, fault, hot spot, etc.), further improving the accuracy and adaptability of the system.

[0156] Generally, the recognition result of an anomaly region is not only a location but can also include the temperature information, fault type, and fault severity of the region. This information will help maintenance personnel more accurately judge the device status and thus take timely and effective maintenance measures.

[0157] As an alternative, in addition to using SVM and decision trees, other advanced deep learning algorithms such as Convolutional Neural Network (CNN) can also be applied to this task, especially in more complex scenarios. By integrating deep learning models, the system can automatically extract richer representations at the feature level, further enhancing the ability to identify abnormal regions.

[0158] After the abnormal region is identified in step S6, the system transmits the identified abnormal region and its location information to the subsequent abnormal region localization and optimization steps. This step further optimizes the detection results of the abnormal region, providing more accurate information for subsequent maintenance decisions.

[0159] S7. Policy Optimization to Improve Recognition Accuracy: Optimize the abnormal region recognition policy through reinforcement learning and global optimization algorithms to improve the accuracy of recognition results;

[0160] Through policy optimization, combining reinforcement learning and global optimization algorithms, the existing recognition policy is optimized, enabling the system to maintain high accuracy and robustness in a changing environment.

[0161] In this embodiment, we adopt reinforcement learning and global optimization algorithms such as genetic algorithms or simulated annealing algorithms to optimize the policy for abnormal region recognition. Through reinforcement learning, the system can adjust the policy according to feedback during continuous trial-and-error processes, and further improve the search scope and stability of the policy through global optimization algorithms, thereby enhancing the accuracy of abnormal region recognition.

[0162] Generally, reinforcement learning is a machine learning method for optimizing policies by interacting with the environment. The system adjusts its behavior by continuously performing actions and receiving reward signals from the environment, with the aim of achieving the maximum long-term reward. In the present invention, the main objective of reinforcement learning is to optimize the recognition policy, enabling the system to more accurately identify and locate abnormal regions.

[0163] Specifically, we define the state space, action space, and reward function of the system:

[0164] State Space: The state space represents the environmental state of the system at each moment. In the present invention, the state space can be the feature information of each pixel point in the image or the overall features of image patches. Each state represents the current abnormal region distribution in the image and other relevant information.

[0165] Action Space: The action space represents different recognition policies that the system can adopt. Each action corresponds to an adjustment of the abnormal region recognition policy, such as selecting different thresholds, adjusting feature weights, etc.

[0166] Reward function: The reward function is used to evaluate whether the actions selected by the system contribute to improving the recognition accuracy. Specifically, if the recognition result is correct, the system will receive a positive reward; if the recognition is incorrect, it will receive a negative reward. In this way, the system gradually learns strategies that can improve the recognition accuracy.

[0167] In a possible implementation, the reinforcement learning method can use Q-learning or Deep Q-Network (DQN) for policy optimization. Q-learning is a value-function-based method that selects the optimal action by learning the Q-values of each state-action pair. The update formula for Q-values is as follows:

[0168] ;

[0169] Where: is the Q-value of taking action in state ; is the reward obtained by the system after taking action ; is the discount factor, which is used to control the influence degree of future rewards; is the learning rate, which is used to control the update speed of Q-values; is the maximum Q-value of taking the best action in state .

[0170] Through this update formula, the Q-values gradually converge, and finally the system can learn the optimal strategy, thereby improving the recognition accuracy of the abnormal area.

[0171] As an option, in addition to Q-learning, we can also adopt Deep Q-Network (DQN). DQN can handle more complex state spaces and high-dimensional data by using a deep neural network to approximate the Q-value function. Through the training of the neural network, the system can be optimized in a large-scale state space, thereby effectively improving the recognition accuracy of the abnormal area.

[0172] In a possible implementation, to further improve the effect of policy optimization, the present invention also introduces a global optimization algorithm, such as a genetic algorithm or a simulated annealing algorithm. These global optimization algorithms optimize the parameters of the policy through global search, avoiding the problem of local optimal solutions, thereby ensuring the global optimality of the recognition policy.

[0173] Genetic algorithm: The genetic algorithm is an optimization algorithm that simulates the process of natural selection. It generates new policy combinations through operations such as selection, crossover, and mutation, and evaluates their advantages and disadvantages according to fitness, gradually evolving to the optimal solution. The main steps of the genetic algorithm include:

[0174] Selection: Select excellent strategies for crossover operation according to the fitness values of each strategy;

[0175] Crossover: Generate new strategies by combining two strategies;

[0176] Mutation: Randomly mutate some strategies to increase diversity;

[0177] Fitness evaluation: Calculate the fitness values of each strategy and evaluate its contribution to the recognition accuracy of abnormal regions.

[0178] Simulated annealing algorithm: The simulated annealing algorithm is a global optimization algorithm based on probabilistic search, which can jump out of the local optimal solution during the search process to find the global optimal solution. Its basic principle is to simulate the physical annealing process, gradually reduce the search range by controlling the temperature, and finally converge to the optimal solution.

[0179] Specifically, by combining reinforcement learning and global optimization algorithms, the present invention can continuously adjust the abnormal region recognition strategy to ensure high recognition accuracy in different scenarios. This optimization process can not only improve the accuracy of the model but also enhance the robustness of the system, enabling it to effectively identify abnormal regions in the face of complex and dynamically changing environments.

[0180] After the strategy optimization, the system will obtain a more accurate and robust abnormal region recognition strategy, which will provide guidance for subsequent abnormal region location and response, ensuring accurate location of abnormal regions and timely response under various environmental conditions.

[0181] The power transmission line abnormal region recognition system with infrared thermal image feature fusion described below can be correspondingly referred to the power transmission line abnormal region recognition method with infrared thermal image feature fusion described above.

[0182] Please refer to the appendix Figure 2 , the present invention also provides a power transmission line abnormal region recognition system with infrared thermal image feature fusion, including:

[0183] A data acquisition module for acquiring infrared thermal image data of the power transmission line;

[0184] A preprocessing module for preprocessing the acquired infrared thermal image data;

[0185] A feature extraction module for performing wavelet transform, multi-scale analysis, data embedding, and nonlinear dynamics modeling on the preprocessed infrared thermal image data to extract multi-dimensional features;

[0186] A graph neural network optimization module for optimizing the graph structure of the extracted features;

[0187] An abnormal area recognition module for locating and recognizing abnormal areas based on the optimized features;

[0188] A reinforcement learning and global optimization module for optimizing the abnormal area recognition strategy and improving the recognition accuracy.

[0189] The system of this embodiment can be used to execute the above method embodiment, and its principle and technical effect are similar, which will not be elaborated here.

[0190] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying abnormal areas of power transmission lines based on infrared thermal image feature fusion, characterized in that: The following steps are involved: Collect infrared thermal imaging data of the transmission line and pre-process the data; The pre-processed infrared thermal image is decomposed into multiple scales by wavelet transform to extract the low-frequency part and the high-frequency part; The extracted multi-scale features are embedded into high-dimensional data through the manifold learning method to obtain a feature set in a low-dimensional space; The spatiotemporal evolution of the abnormal area is modeled based on a nonlinear dynamic model and fused with the embedded feature set; Use graph neural networks to optimize the fused features, update the relationship between nodes, and output the optimized features; Use the optimized features to identify abnormal areas and obtain the locations of abnormal areas; The abnormal area recognition strategy is optimized through reinforcement learning and global optimization algorithms to improve the accuracy of recognition results. The nonlinear dynamic model adopts the Lorenz system to describe the dynamic evolution process of the abnormal area. The dynamic evolution process includes: Based on the nonlinear equation of the Lorenz system, the dynamic evolution of abnormal areas in infrared thermal images over time is simulated. By using the three-dimensional state variables of the Lorenz equation, the temperature change, expansion and shrinkage of the abnormal area are described. In the Lorenz system, by selecting appropriate parameters and adjusting according to the actual infrared thermal imaging data, the model can better match the actual temperature change pattern; The Lorenz system consists of three equations and is used to describe a three-dimensional nonlinear system. Its basic form is as follows: ； ； ； in: , , Represent the three state variables of the Lorenz system, represents the temperature change, represents the speed at which the hot spot expands. Represents other relevant temperature factors or equipment status; , , are the parameters of the Lorenz system, set as the initial conditions of the system; The dynamic evolution process obtained by Lorenz system modeling is fused with the multi-scale features extracted from infrared thermal images through a weighted fusion method.

2. The method for identifying abnormal areas of power transmission lines by fusion of infrared thermal image features according to claim 1 is characterized in that: The pre-processing comprises: Remove noise from the collected infrared thermal imaging data; Perform contrast enhancement processing on infrared thermal images; Perform histogram equalization on the processed image.

3. The method for identifying abnormal areas of power transmission lines by fusion of infrared thermal image features according to claim 1 is characterized in that: The wavelet transform adopts discrete wavelet transform, the low-frequency part includes the overall temperature distribution characteristics, the high-frequency part includes the detail characteristics, and the extraction process of the low-frequency part and the high-frequency part includes: Perform wavelet transform on the image data to obtain low-frequency and high-frequency features at different scales; The low-frequency and high-frequency parts are denoised, and the parts with significant features are retained.

4. The method for identifying abnormal areas of power transmission lines by fusion of infrared thermal image features according to claim 1 is characterized in that: The high-dimensional data embedding is performed using a principal component analysis method or a t-SNE method, and the feature set after the data embedding includes: Perform dimensionality reduction on high-dimensional data to extract the most representative features; The reduced feature set is used as input for subsequent processing.

5. The method for identifying abnormal areas of power transmission lines by fusion of infrared thermal image features according to claim 1, characterized in that: The graph neural network optimization includes: Each image feature node is used as a node of the graph, and the similarity between nodes is represented by the edges of the graph; Normalize the adjacency matrix of the graph, use graph neural network to propagate features, and update node features; The node features are integrated using graph convolution operations to obtain the optimized feature set as the final output.

6. The method for identifying abnormal areas of power transmission lines by fusion of infrared thermal image features according to claim 1, characterized in that: The abnormal area recognition adopts support vector machine or decision tree method, and the recognition results include: Establish a classification model based on training data to automatically identify abnormal areas; Output the location coordinates and temperature information of the identified abnormal area.

7. The method for identifying abnormal areas of power transmission lines by fusion of infrared thermal image features according to claim 1, characterized in that: The reinforcement learning algorithm includes: Define state space, action space, and reward function; Use Q-learning or deep Q-network to optimize abnormal region identification strategy; The strategy is adjusted according to the feedback reward function to maximize the recognition accuracy.

8. The method for identifying abnormal areas of power transmission lines by fusion of infrared thermal image features according to claim 1, characterized in that: The global optimization uses a genetic algorithm or a simulated annealing algorithm to globally optimize the identification strategy, and the optimization process includes: Searching for parameters of the identification strategy using genetic algorithms or simulated annealing algorithms; The stability and accuracy of the strategy are optimized through global search.

9. The abnormal area recognition system of power transmission line based on infrared thermal image feature fusion is characterized by: The method for identifying abnormal areas of power transmission lines using the infrared thermal image feature fusion described in any one of claims 1 to 8 comprises: Data acquisition module, used to collect infrared thermal imaging data of power transmission lines; A preprocessing module, used for preprocessing the collected infrared thermal imaging data; Feature extraction module, used to perform wavelet transform, multi-scale analysis, data embedding and nonlinear dynamic modeling on the pre-processed infrared thermal imaging data to extract multi-dimensional features; Graph neural network optimization module, used to optimize the graph structure of the extracted features; An abnormal area recognition module is used to locate and identify abnormal areas based on optimized features; Reinforcement learning and global optimization modules are used to optimize abnormal area identification strategies and improve identification accuracy.

Images