Power transmission line typical component thermal hidden danger identification method, device, equipment and medium

Abstract

The invention discloses a thermal hidden danger identification method, device and equipment for typical parts of a power transmission line and a medium, and belongs to the technical field of power system monitoring, and the method comprises the steps: constructing a multi-modal image segmentation model, and segmenting an infrared image containing the typical parts of the power transmission line through the model; based on the segmentation result, gray values of the typical parts are extracted, temperature values corresponding to the gray values are obtained in combination with preset temperature information, and temperature gray value data pairs are formed; performing anomaly detection on the data pair, identifying and removing abnormal data points, and obtaining cleaned temperature gray value data; and performing B spline curve fitting on the cleaned temperature gray value data, establishing a curvilinear relationship between the temperature and the gray value, and identifying the thermal hidden danger of the typical component of the power transmission line based on the curvilinear relationship. According to the method, the accuracy and robustness of segmentation of the typical parts of the power transmission line are improved, the accuracy of the temperature gray mapping relation is improved, and therefore more reliable thermal hidden danger recognition is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring, and in particular to a method, device, equipment and medium for identifying thermal hazards of typical components of a transmission line. Background Art

[0002] Infrared thermal imaging technology is widely used in electrical equipment condition monitoring due to its non-contact, real-time monitoring capabilities. It can effectively detect local thermal anomalies in equipment. Existing methods for diagnosing thermal faults in electrical equipment fall into two main categories: those based on traditional image feature analysis and those based on deep neural networks.

[0003] However, research on thermal fault diagnosis of transmission line equipment primarily relies on predefined thermal region segmentation or detection rules to achieve fault identification. Although infrared images inherently contain temperature information, in practice, due to interference from various external factors such as ambient temperature fluctuations and climate change, relying solely on image data often makes it difficult to accurately determine thermal faults. Furthermore, when extracting pixel and temperature data, existing methods struggle to accurately identify the actual area of ​​the equipment, affecting the accuracy of the temperature-to-grayscale mapping relationship and, in turn, the diagnostic effectiveness. Finally, the data extraction process may also incorporate anomalous data points unrelated to typical transmission line components, such as those caused by environmental interference. This fails to fully consider data validity, leading to deviations in the established temperature-to-grayscale mapping relationship and affecting the accuracy and reliability of thermal hazard identification. Summary of the Invention

[0004] Based on this, it is necessary to propose a method, device, equipment and medium for identifying thermal hazards of typical components of transmission lines to address the above problems.

[0005] The present invention provides a method for identifying thermal hazards of typical components of a power transmission line, the method comprising:

[0006] Constructing a multimodal image segmentation model, and using the multimodal image segmentation model to segment an infrared image containing typical components of a power transmission line, thereby obtaining a segmentation result of the typical components of the power transmission line, wherein the multimodal image includes an infrared image;

[0007] Based on the segmentation result, extracting the grayscale value of the typical component in the infrared image, and obtaining the temperature value corresponding to the grayscale value in combination with the preset temperature information to form a temperature-grayscale value data pair;

[0008] Performing anomaly detection on the temperature grayscale value data pairs, identifying and eliminating abnormal data points, and obtaining cleaned temperature grayscale value data;

[0009] B-spline curve fitting is performed on the temperature grayscale value data after cleaning to establish a curve relationship between temperature and grayscale value, and based on the curve relationship, the temperature value of typical components is calculated to identify thermal hazards of typical components of the transmission line.

[0010] In some embodiments, the multimodal image segmentation model is obtained based on the Transformer architecture, multimodal information and transfer learning strategy; the multimodal image segmentation model includes multi-head self-attention, a feedforward neural network and a classification head.

[0011] In some embodiments, constructing a multimodal image segmentation model and using the multimodal image segmentation model to segment an infrared image containing typical components of a transmission line to obtain a segmentation result of the typical components of the transmission line includes:

[0012] The input infrared image is divided into patches of fixed size, each patch is processed, flattened into a one-dimensional vector, and then converted into a vector of dimension D through linear mapping. The patches are arranged in order to form a sequence, and a learnable classification tag and position encoding are introduced into the sequence to form a final sequence;

[0013] Fusing the visible light image and the infrared image, and inputting the obtained fused image into a feature extraction network to obtain an intermediate feature map F;

[0014] Applying global maximum pooling to the intermediate feature map F and mapping it into prompt information through a feedforward network, and inputting the prompt information into the mask encoding module of infrared image segmentation;

[0015] Encoding the final sequence through a multi-head self-attention and feedforward neural network, and combining the prompt information of the fused image branch to obtain the final encoder output;

[0016] The final encoder output is classified by the classification head to obtain the segmentation result of the infrared image containing typical components of the transmission line.

[0017] In some embodiments, the preset temperature information includes a preset mapping relationship or calibration curve between grayscale values ​​and temperature;

[0018] The grayscale value is extracted from the original infrared image based on the segmentation result, and the corresponding temperature value is obtained in combination with the preset temperature information to form a temperature grayscale value data pair, including:

[0019] Determining pixel locations of typical components of the power transmission line based on the segmentation results;

[0020] Extracting the grayscale value at the pixel position from the infrared image, and determining temperature data corresponding to the grayscale value at the pixel position according to a preset mapping relationship between grayscale value and temperature or a calibration curve;

[0021] The grayscale value is combined with the corresponding temperature data to form the temperature grayscale value data pair.

[0022] In some embodiments, performing anomaly detection on the temperature grayscale value data, identifying and removing abnormal data points, and obtaining cleaned temperature grayscale value data includes:

[0023] Using the isolation forest method to detect the temperature grayscale value data pair, obtain a first anomaly indicator, set a first anomaly score threshold, and screen out a candidate subset containing anomalies that meet the first anomaly score threshold;

[0024] The candidate subset is tested again using the local anomaly factor method to obtain a second anomaly index.

[0025] Assigning a comprehensive anomaly score to each data point in the temperature grayscale value data pair by combining the first anomaly indicator and the second anomaly indicator;

[0026] A second scoring threshold is set according to the comprehensive abnormality score, and abnormal points in the temperature grayscale value data pair are determined and eliminated according to the second scoring threshold to obtain cleaned temperature grayscale value data.

[0027] In some embodiments, determining and removing abnormal points in the temperature grayscale value data pair according to the second scoring threshold includes:

[0028] After determining the abnormal points in the temperature grayscale value data pair according to the second scoring threshold, calculating the abnormal proportion;

[0029] If the abnormal ratio does not exceed the preset abnormal ratio threshold, the abnormal points in the temperature gray value data pair are removed;

[0030] If the abnormality ratio exceeds the preset abnormality ratio threshold, the second abnormality scoring threshold is lowered, and the abnormal points in the temperature grayscale value data pair are re-determined and the abnormality ratio is calculated based on the lowered second scoring threshold. When the abnormality ratio does not exceed the preset abnormality ratio threshold, the abnormal points in the temperature grayscale value data pair are eliminated.

[0031] In some embodiments, performing B-spline curve fitting on the temperature grayscale value data after cleaning includes:

[0032] Determining the number of nodes, wherein the number of nodes refers to the number of uniform nodes or the number of non-uniform nodes;

[0033] Determine the B-spline degree;

[0034] The control points of the temperature grayscale value data are calculated by the least square method, and the curve relationship between the temperature and the grayscale value is determined based on the control points.

[0035] The present application also provides a device for identifying thermal hazards of typical components of a power transmission line, the device comprising:

[0036] a model building module, configured to build a multimodal image segmentation model, and use the multimodal image segmentation model to segment infrared images containing typical components of a transmission line, thereby obtaining segmentation results of the typical components of the transmission line;

[0037] a data processing module, configured to extract the grayscale value of a typical component from the infrared image, and obtain the temperature value corresponding to the grayscale value in combination with preset temperature information, thereby forming a temperature-grayscale value data pair;

[0038] An anomaly detection module is used to perform anomaly detection on the temperature gray value data pair, identify and eliminate abnormal data points, and obtain cleaned temperature gray value data;

[0039] The thermal hazard identification module is used to perform B-spline curve fitting on the temperature grayscale value data after cleaning, establish a curve relationship between temperature and grayscale value, and calculate the temperature value of typical components based on the curve relationship to identify thermal hazards of typical components of the transmission line.

[0040] An embodiment of the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0041] Constructing a multimodal image segmentation model, and using the multimodal image segmentation model to segment a multimodal infrared image containing typical components of a power transmission line, to obtain a segmentation result of the typical components of the power transmission line, wherein the multimodal image includes an infrared image;

[0042] Based on the segmentation result, extracting the grayscale value of the typical component in the infrared image, and obtaining the temperature value corresponding to the grayscale value in combination with the preset temperature information to form a temperature-grayscale value data pair;

[0043] Performing anomaly detection on the temperature grayscale value data pairs, identifying and eliminating abnormal data points, and obtaining cleaned temperature grayscale value data;

[0044] B-spline curve fitting is performed on the temperature grayscale value data after cleaning to establish a curve relationship between temperature and grayscale value, and based on the curve relationship, the temperature value of typical components is calculated to identify thermal hazards of typical components of the transmission line.

[0045] The present application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor performs the following steps:

[0046] Constructing a multimodal image segmentation model, and using the multimodal image segmentation model to segment an infrared image containing typical components of a power transmission line, thereby obtaining a segmentation result of the typical components of the power transmission line, wherein the multimodal image includes an infrared image;

[0047] Based on the segmentation result, extracting the grayscale value of the typical component in the infrared image, and obtaining the temperature value corresponding to the grayscale value in combination with the preset temperature information to form a temperature-grayscale value data pair;

[0048] Performing anomaly detection on the temperature grayscale value data pairs, identifying and eliminating abnormal data points, and obtaining cleaned temperature grayscale value data;

[0049] B-spline curve fitting is performed on the temperature grayscale value data after cleaning to establish a curve relationship between temperature and grayscale value, and based on the curve relationship, the temperature value of typical components is calculated to identify thermal hazards of typical components of the transmission line.

[0050] The embodiments of the present application have the following beneficial effects:

[0051] In the method for identifying thermal hazards of typical components of a transmission line provided in an embodiment of the present application, a multimodal image segmentation model is constructed, and the multimodal image containing the typical components of the transmission line is segmented using the multimodal image segmentation model; based on the segmentation results, the grayscale values ​​of the typical components are extracted, and the temperature values ​​corresponding to the grayscale values ​​are obtained in combination with preset temperature information to form temperature grayscale value data pairs; anomaly detection is performed on the temperature grayscale value data pairs, abnormal data points are identified and eliminated, and cleaned temperature grayscale value data is obtained; B-spline curve fitting is performed on the cleaned temperature grayscale value data to establish a curve relationship between temperature and grayscale value, and thermal hazards of the typical components of the transmission line are identified based on the curve relationship. This helps to solve the problems in the prior art of relying solely on infrared image data to accurately identify thermal hazards in complex environments, and the problem of insufficient accuracy of the temperature grayscale mapping relationship due to the failure to accurately segment equipment areas and effectively eliminate abnormal data points. It realizes efficient identification and monitoring of thermal hazards of transmission lines, thereby discovering and warning of potential equipment failure risks in advance, and ensuring the safety and stability of power grid operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0053] in:

[0054] Figure 1 1. A schematic flow chart of a method for identifying thermal hazards of typical components of a power transmission line according to an embodiment;

[0055] Figure 2 Schematic diagram of infrared image segmentation by a multimodal image segmentation model in one embodiment;

[0056] Figure 3 1. A diagram showing the experimental results of segmenting typical transmission line components on a test dataset in one embodiment;

[0057] Figure 4 Graph showing data corresponding to temperature grayscale values ​​and B-spline curve fitting results in one embodiment;

[0058] Figure 5 A structural diagram of a device for identifying thermal hazards of typical components of a power transmission line in one embodiment;

[0059] Figure 6 is a schematic structural diagram of a computer device in one embodiment;

[0060] Figure 7 FIG. 1 is a schematic diagram of the structure of a computer-readable storage medium in one embodiment. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0062] First, the key terms of this application are explained as follows:

[0063] Transformer architecture: The Transformer is a deep learning model architecture based on the self-attention mechanism, widely used in natural language processing and computer vision. It can efficiently capture global dependencies in sequential data and has strong parallel computing capabilities.

[0064] Self-attention mechanism: The self-attention mechanism is a technique that generates weighted representations by calculating the similarity between each element in the input sequence. It helps the model focus on important information related to the current task when processing long sequences of data, reducing information loss.

[0065] Image segmentation: Image segmentation is a fundamental task in computer vision. Its goal is to divide an image into several regions of similarity or consistency, thereby distinguishing different objects from the background. This task has important applications in medical image analysis, autonomous driving, remote sensing monitoring, video surveillance, and other fields. Existing image segmentation research methods are mainly divided into those based on traditional image feature analysis and those based on deep neural networks.

[0066] In the embodiment of the present application, a method for identifying thermal hazards of typical components of a transmission line is provided. Figure 1 , Figure 1 1 is a flow chart of a method for identifying thermal hazards of typical components of a transmission line in one embodiment; the method for identifying thermal hazards of typical components of a transmission line includes steps S1 to S4.

[0067] Step S1, constructing a multimodal image segmentation model, and using the multimodal image segmentation model to segment an infrared image containing typical components of a transmission line, to obtain a segmentation result of the typical components of the transmission line;

[0068] In some embodiments, the multimodal image segmentation model is obtained based on the Transformer architecture, multimodal information and transfer learning strategy; the multimodal image segmentation model includes multi-head self-attention, a feedforward neural network and a classification head.

[0069] Specifically, the multimodal image segmentation model, unlike the original image shake correction model, aims to accurately segment typical components of power transmission lines (such as conductors and insulators). This model fully accounts for potential noise, blur, and environmental interference in infrared images, integrating transfer learning and multimodal information. Leveraging the global information capture capabilities of the Transformer architecture, it automatically learns and extracts features of the target region. Ultimately, this multimodal image segmentation model effectively separates the target region from the background, laying a solid foundation for subsequent extraction of temperature and grayscale value data.

[0070] In some embodiments, constructing a multimodal image segmentation model and using the multimodal image segmentation model to segment an infrared image containing typical components of a transmission line to obtain a segmentation result of the typical components of the transmission line includes:

[0071] The input infrared image is divided into patches of fixed size, each patch is processed, flattened into a one-dimensional vector, and then converted into a vector of dimension D through linear mapping. The patches are arranged in order to form a sequence, and a learnable classification tag and position encoding are introduced into the sequence to form a final sequence;

[0072] Fusing the visible light image and the infrared image, and inputting the obtained fused image into a feature extraction network to obtain an intermediate feature map F;

[0073] Applying global maximum pooling to the intermediate feature map F and mapping it into prompt information through a feedforward network, and inputting the prompt information into the mask encoding module of infrared image segmentation;

[0074] Encoding the final sequence through a multi-head self-attention and feedforward neural network, and combining the prompt information of the fused image branch to obtain the final encoder output;

[0075] The final encoder output is classified by the classification head to obtain the segmentation result of the infrared image containing typical components of the transmission line.

[0076] Specifically, if Figure 2 In one embodiment shown, a multimodal image segmentation model is used to segment infrared images. Assume that the input image is Where H and W are the height and width of the image, and C is the number of channels. The image is divided into fixed-size patches, each of which is P×P in size. For each patch. First flatten it into a one-dimensional vector: Then transform it into a vector of dimension D through linear mapping In this way, each patch is converted into a D-dimensional vector.

[0077] Arrange the embedding vectors of all patches in order to form a sequence: Z patch =[z p1 ,z p2 ,…,z pN ], and introduce a learnable classification tag Add it to the front of the sequence and add the position code Z0=[z cls ;z p1 ,z p2,…,z pN ]+E pos , where the encoding of each position ensures that the model can utilize the position information of each patch in the image.

[0078] Transformer is composed of multiple layers with the same structure. Each layer contains two main modules: Multi-head Self-Attention (MSA) and Feedforward Neural Network (MLP), while using residual connection and layer normalization (LN). For the lth layer, let the input be z (l-1) , first perform layer normalization on the input:

[0079]

[0080] The calculation formula of layer normalization is:

[0081]

[0082] Where μ and σ are the mean and standard deviation of the input x, respectively, and γ and β are learnable parameters.

[0083] Next, calculate Q, K, and V:

[0084]

[0085] in is the parameter matrix and d is the dimension of each attention head.

[0086] The attention of a single head is calculated as:

[0087]

[0088] If there are h heads, the output of Multi-Head Self-Attention (MSA) is:

[0089]

[0090] For the i-th head:

[0091]

[0092] in is the output projection matrix.

[0093] Then add the residual connection to get the output of the multi-head self-attention module:

[0094]

[0095] Perform layer normalization again on the output of the previous step:

[0096]

[0097] Then, an MLP consisting of two layers of full connections and activation functions (such as GELU) is constructed:

[0098]

[0099] in And b1, b2 are learnable parameters.

[0100] Similarly, adding residual connections finally gives the output of the current layer:

[0101]

[0102] Repeat the above process to stack L layers to obtain the final encoder output z(L).

[0103] In the output sequence of the Transformer encoder, the first token corresponds to the CLS token, which is finally represented as Use this representation for classification:

[0104]

[0105] in and b head is the parameter of the classification head, and K is the number of categories.

[0106] The convolution layer uses multiple filters to calculate the feature representation of the input visual image to achieve feature space mapping. The convolution layer parameters include the convolution kernel size, edge padding, and stride length, which have a direct effect on the image feature output. The bottom convolution layer is mainly responsible for learning the underlying features such as the edge curve of the visual image, and the high-level convolution layer is responsible for learning the abstract features of the visual image. Taking a visual image of size w×h as input, the size of the visual image feature map obtained after n convolution kernels with size k, edge padding value p, and stride length s is defined as w1×h1×n. The calculation formulas for w1 and h1 are as follows:

[0107] w1=(w+2 * pk) / s+1(12)

[0108] h1=(h+2 * pk) / s+1(13)

[0109] The pooling layer uses downsampling operations to reduce the dimensionality of visual image features obtained after continuous convolutional layers to achieve a feature information filtering effect. Common pooling methods include maximum pooling and average pooling. Visual images processed by convolutional pooling can maintain feature, rotation, and translation invariance while reducing the number of CNN network model parameters and enhancing model fitting capabilities. Define the convolutional pooling layer step size as s, the pre-parameter as p, and the pixel point as (i, j). The pooling operation calculation formula is shown below:

[0110]

[0111] It can be seen that when the pre-parameter p tends to infinity, the formula is expressed as maximum pooling; when p is equal to 1, the formula is expressed as mean pooling. The output characteristics of the pooling operation are jointly affected by the stride length, edge padding, and the size of the pooling parameter.

[0112] Assumptions is the final output, where N p To represent the number of cue points, we first need to map these low-dimensional cues into a high-dimensional space through an MLP. The formula is:

[0113] E p =MLP(P)=σ(W2(σ(W1P+b1))+b2)(15)

[0114] In order to effectively integrate the prompt information with the image context features, a cross-attention mechanism is adopted. The specific steps for calculating Q, K, and V are as follows:

[0115] Q is embedded in E by the prompt p generate

[0116]

[0117] K and V are generated by image features F

[0118] K=FW K ,V=FW V (17)

[0119] Calculate the attention weight using Q and K to calculate the attention score, and get the weight through softmax:

[0120]

[0121] Perform weighted summation of the value V using the attention weight A:

[0122] E cross =AV,

[0123] Combine the output of the cross-attention with the initial hint embedding, typically using residual connections and layer normalization:

[0124] E prompt =LayerNorm(E p +E cross ) (20)

[0125] This can effectively stabilize the training and preserve the original prompt information.

[0126] To further enhance the expressive power of the hint embedding, the above results can be input into a feedforward network, and residual connections and normalization are performed again:

[0127] E′ prompt =LayerNorm(E prompt +FFN(E prompt )) (twenty one)

[0128] Among them, FFN(·) usually consists of two layers of fully connected networks and nonlinear activation functions.

[0129] The goal of this application is to accurately identify typical components of power transmission lines in infrared images. Considering the small number of infrared image datasets, transfer learning is used to improve the performance of the model. Transfer learning is an important branch of machine learning. Since the 1990s, many related algorithms have been proposed. Traditional machine learning methods train models independently for different tasks, while transfer learning, based on existing models, identifies the similarities between source domain and target domain tasks and transfers existing knowledge to the target domain to speed up the adaptation of new tasks. In order to ensure the effectiveness of the target model, transfer learning usually requires a certain correlation between the source domain and the target domain, such as the similarity of data or tasks. The higher the similarity between the source domain and the target domain, the better the effect of transfer learning.

[0130] Transfer learning mainly includes domains and tasks. The domain D consists of the feature space X and its marginal probability distribution P(X), that is, D = {X, P(X)}. The task T is composed of the label space Y and the prediction function f(X), that is, T = {Y, f(X)}. Transfer learning is achieved by transferring the source domain D to the target domain. s ={X s ,P s (X)} to the target domain D T ={X T ,P T (X)}, which can improve the performance of the target domain task and plays a greater role when the target domain dataset is small.

[0131] Deep transfer learning is a widely used method in transfer learning. It pre-trains a neural network on a large dataset and adjusts the parameters of some layers for the target task. Fine-tuning the parameters of a deep learning model trained on a large number of source domain samples is performed on a small number of target domain samples. Leveraging the deep learning model's excellent generalization properties, this reduces the training time and sample size, while improving the performance of the target domain task.

[0132] The basic model used in this study is Segmentation Anything (SeA), which is built on the Transformer architecture. The self-attention mechanism in the Transformer architecture can capture long-range dependencies in sequential data. In image segmentation, the feature map is flattened into a sequence representation, and the feature vector at each position is treated as an element in the sequence and input into the Transformer encoder.

[0133] Self-attention mechanism: The core of the self-attention mechanism is to focus on each element by adjusting the degree of association between each element (such as attention weight). This study calculates attention weights and sums them, and the self-attention mechanism can dynamically aggregate feature information at different locations in the image.

[0134] Multi-head attention mechanism: To better capture different patterns in features, a multi-head attention mechanism is introduced. The input feature sequence is divided into multiple subspaces, and self-attention operations are performed on each subspace. This multi-head attention mechanism allows the model to learn different patterns from different feature subspaces, further improving segmentation performance.

[0135] Position Encoding: The Transformer architecture itself does not contain position information, so position encoding is added to the input sequence to capture the spatial relationship in the image.

[0136] The prompt encoder is a hint encoder. The embedding vectors generated by the hint encoder provide clear instructions to the mask decoder, enabling the segmentation network to more accurately locate and extract the target area, maintaining high segmentation accuracy even in complex backgrounds or with significant interference. Because hint information is generally sparse, its encoding process does not require a very deep network structure, thereby ensuring overall computational efficiency while avoiding the introduction of too many parameters that may lead to unstable model training. The hint encoder uses different encoding strategies for different hint types, mainly including: point hints, bounding box hints, and text hints.

[0137] Point cues typically consist of a series of two-dimensional coordinates, with each point labeled as positive (belonging to the target) or negative (not belonging to the target). First, the two-dimensional coordinates are mapped to a high-dimensional feature space using a linear layer or a multi-layer perceptron (MLP). Position encoding is also added to preserve the point's position information, ensuring that the encoded vector contains not only the category information but also the spatial location information.

[0138] The bounding box is usually represented by the coordinates of the upper left and lower right corners or the center point and the width and height. The bounding box parameters are mapped to a fixed-dimensional embedding vector through a linear transformation. This encoding method can capture the global position and size information of the bounding box, providing an approximate area range for subsequent segmentation. In some applications, a preliminary mask or text description may be input. This part is usually handled by a specially designed encoding module (such as a lightweight convolutional network or a pre-trained text encoder), but in the basic design of SeA, points and boxes are mainly used.

[0139] Step S2, based on the segmentation result, extracting the grayscale value of the typical component in the infrared image, and obtaining the temperature value corresponding to the grayscale value in combination with preset temperature information to form a temperature-grayscale value data pair;

[0140] Specifically, after image segmentation, data extraction is performed on the segmented target areas. Specifically, the corresponding temperature information and grayscale value data are extracted from each target area. These data reflect the local thermal distribution and grayscale performance of the device. Through precise segmentation and data extraction, a corresponding relationship between the temperature and grayscale value of each component can be established, that is, a temperature-grayscale value data pair. This captures local temperature anomalies that may exist during device operation and provides detailed data information for thermal hazard detection.

[0141] In some embodiments, the preset temperature information includes a preset mapping relationship or calibration curve between grayscale values ​​and temperature;

[0142] The grayscale value is extracted from the original infrared image based on the segmentation result, and the corresponding temperature value is obtained in combination with the preset temperature information to form a temperature grayscale value data pair, including:

[0143] Determining pixel locations of typical components of the power transmission line based on the segmentation results;

[0144] Extracting the grayscale value at the pixel position from the infrared image, and determining temperature data corresponding to the grayscale value at the pixel position according to a preset mapping relationship between grayscale value and temperature or a calibration curve;

[0145] The grayscale value is combined with the corresponding temperature data to form the temperature grayscale value data pair.

[0146] Specifically, after image segmentation, the generated mask is first used to determine the pixel locations in the target area, and the corresponding pixel coordinates are then located in the original image. Next, based on the grayscale values ​​of these pixels, a match is performed within a pre-defined temperature coordinate area on the right side of the image. This area typically contains a temperature scale or calibrated temperature information. Once the match is complete, the corresponding temperature data is calculated using the known mapping relationship between grayscale values ​​and temperature or a calibration curve. Finally, the extracted temperature data is combined with the original grayscale values ​​to form a temperature-grayscale value data pair, providing an accurate data foundation for subsequent data analysis and processing.

[0147] Step S3, performing anomaly detection on the temperature grayscale value data pair, identifying and removing abnormal data points, and obtaining cleaned temperature grayscale value data;

[0148] Specifically, the Local Outlier Factor (LOF) is a density-based anomaly detection method proposed by Breunig et al. in 2000. The core idea of ​​this method is to determine whether a data point is an outlier by comparing its local density. If the density of a point is significantly lower than the density of its neighbors, the point is considered an outlier. This method is suitable for detecting local outliers and can adapt to datasets with different density distributions. However, it has high computational complexity and is computationally expensive for large datasets. In addition, the k value must be manually set.

[0149] Isolation Forest (iForest) is a tree-based anomaly detection method proposed by Liu et al. in 2008. This method is based on the assumption that outliers are more easily isolated, namely, that they are often sparse and have a different feature distribution from normal points. Using a randomized binary partitioning method, outliers are more easily segmented than normal points, so the average depth of the isolation tree can be used as a measure of the anomaly score. This method is computationally efficient, applicable to large datasets, independent of data density or distribution, and does not require data normalization. However, it is important to select an appropriate number of trees, T. Too few trees may result in unstable detection, while too many trees increase the computational load. Furthermore, this method is suitable for global anomaly detection.

[0150] To improve anomaly detection capabilities, we combined the strengths of the two aforementioned methods to employ a dynamic fusion LOF-IF model to detect and remove outliers from the data. LOF excels at detecting local anomalies, but suffers from high computational complexity; IF, on the other hand, excels at global anomaly detection and offers high computational efficiency. Combining the two approaches allows for complementary advantages.

[0151] In some embodiments, performing anomaly detection on the temperature grayscale value data, identifying and removing abnormal data points, and obtaining cleaned temperature grayscale value data includes:

[0152] Using the isolation forest method to detect the temperature grayscale value data pair, obtain a first anomaly indicator, set a first anomaly score threshold, and screen out a candidate subset containing anomalies that meet the first anomaly score threshold;

[0153] The candidate subset is tested again using the local anomaly factor method to obtain a second anomaly index.

[0154] Assigning a comprehensive anomaly score to each data point in the temperature grayscale value data pair by combining the first anomaly indicator and the second anomaly indicator;

[0155] A second scoring threshold is set according to the comprehensive abnormality score, and abnormal points in the temperature grayscale value data pair are determined and eliminated according to the second scoring threshold to obtain cleaned temperature grayscale value data.

[0156] Specifically, the first method of fusing LOF and IF in the dynamic fusion LOF-IF model is:

[0157] Initial screening: First, Isolation Forest is applied to perform a preliminary screening of the dataset, that is, to detect the temperature grayscale value data pairs. Since IF has low time complexity, it can quickly process large-scale data. By setting a low first anomaly score threshold, candidate subsets that may contain outliers are screened out. At the same time, a higher threshold for extreme anomaly scores is set to eliminate obvious outliers.

[0158] Fine detection: LOF is applied to the candidate subset obtained by preliminary screening for fine detection. LOF can effectively identify local anomalies by comparing the local density difference between the data point and its neighborhood.

[0159] Comprehensive evaluation: Combining the detection results of IF and LOF, a weighted average strategy is adopted to integrate the results of the two algorithms and assign a comprehensive anomaly score to each data point.

[0160] Identify outliers: Based on the comprehensive anomaly score, set the second scoring threshold to ultimately identify outliers in the dataset.

[0161] The second method of fusing LOF and IF in the dynamic fusion LOF-IF model is:

[0162] In some embodiments, performing anomaly detection on the temperature grayscale value data, identifying and removing abnormal data points, and obtaining cleaned temperature grayscale value data includes:

[0163] Using the isolation forest method to detect the temperature gray value data pair to obtain a first abnormality indicator;

[0164] Detecting the temperature grayscale value data pair using a local anomaly factor method to obtain a second anomaly indicator;

[0165] Assigning a comprehensive anomaly score to each data point in the temperature grayscale value data pair in combination with the first anomaly indicator and the second anomaly indicator;

[0166] A second scoring threshold is set according to the comprehensive abnormality score, and abnormal points in the temperature grayscale value data pair are determined and eliminated according to the second scoring threshold to obtain cleaned temperature grayscale value data.

[0167] Among them, the first method of fusing LOF and IF has a screening subset, which processes less data and is more efficient.

[0168] Furthermore, based on the two fusion methods, the dynamic fusion LOF-IF model can also be dynamically adjusted.

[0169] In some embodiments, determining and removing abnormal points in the temperature grayscale value data pair according to the second scoring threshold includes:

[0170] After determining the abnormal points in the temperature grayscale value data pair according to the second scoring threshold, calculating the abnormal proportion;

[0171] If the abnormal ratio does not exceed the preset abnormal ratio threshold, the abnormal points in the temperature gray value data pair are removed;

[0172] If the abnormality ratio exceeds the preset abnormality ratio threshold, the second abnormality scoring threshold is lowered, and the abnormal points in the temperature grayscale value data pair are re-determined and the abnormality ratio is calculated based on the lowered second scoring threshold. When the abnormality ratio does not exceed the preset abnormality ratio threshold, the abnormal points in the temperature grayscale value data pair are eliminated.

[0173] Specifically, after detection, the actual anomaly ratio is calculated. If there are too many anomalies, the second anomaly score threshold is automatically lowered and LOF is used again for fine screening.

[0174] Step S4, performing B-spline curve fitting on the temperature grayscale value data after cleaning, establishing a curve relationship between temperature and grayscale value, and identifying thermal hazards of typical components of the power transmission line based on the curve relationship;

[0175] Specifically, after completing anomaly detection, further curve fitting analysis is performed on the temperature grayscale value data. The smoothness and flexibility of the B-spline function are used to fit the data to obtain a continuous curve between temperature and grayscale value. This curve can not only intuitively reflect the overall change trend of the data, but also reveal local details, providing an accurate model for the quantitative analysis of thermal hazards in transmission lines. Through curve fitting, we can better understand the thermal distribution characteristics of equipment in different states, and thus provide scientific support for fault warning and maintenance decisions.

[0176] In some embodiments, performing B-spline curve fitting on the temperature grayscale value data after cleaning includes:

[0177] Determining the number of nodes, wherein the number of nodes refers to the number of uniform nodes or the number of non-uniform nodes;

[0178] Determine the B-spline degree;

[0179] The control points of the temperature grayscale value data are calculated by the least square method, and the curve relationship between the temperature and the grayscale value is determined based on the control points.

[0180] Specifically, B-spline (Basis Spline) is an extended form of spline function, composed of piecewise polynomials, and describes complex curves through linear combinations of basis functions. Compared with traditional polynomial fitting and Bezier curves, B-spline has local support and flexible controllability, and is particularly suitable for processing non-uniformly distributed data and high-noise scenarios. B-spline has the following characteristics: each basis function is non-zero only in a local interval, and modifying the control point only affects the local curve shape; it can define the connection smoothness of different intervals through node vectors; it supports non-uniform node distribution and adapts to changes in data density.

[0181] The definition of B-spline basis function is: given a node vector U = {u0,u1,...,u m}, the i-th p-order spline basis function is defined as:

[0182] Zeroth-order basis function (p=0)

[0183]

[0184] High-order basis functions (p>0)

[0185]

[0186] The B-spline curve consists of n+1 control points P i And basis function combination:

[0187]

[0188] The steps for using the B-spline function to fit data are as follows:

[0189] Determine the number of nodes: The number of nodes can be set in two ways: uniform nodes and non-uniform nodes. Uniform nodes are simple but may ignore the data distribution characteristics. Non-uniform nodes are set based on the data density distribution. Common methods include chord length parameterization and centripetal parameterization.

[0190] Select the B-spline degree: low degree gives priority to smoothness and strong noise resistance. High degree fits complex curves but is prone to overfitting;

[0191] Calculate control points: solve the control point P by the least squares method i .

[0192]

[0193] Fit evaluation and optimization: perform residual analysis to check the distribution of fitting errors, then use cross-validation to split the data set to verify generalization ability and suppress overfitting by adding smoothing terms;

[0194]

[0195] This application introduces curve fitting technology through in-depth analysis of the temperature and grayscale value data extracted after segmentation, aiming to establish a stable mathematical relationship between temperature and image grayscale value. By fitting the temperature grayscale data using the B-spline function, a smooth curve that highly fits the changing trend of the real data can be obtained. The core advantage of this method is that it does not directly rely on image recognition to judge thermal hazards, but instead accurately calculates the actual temperature values ​​of typical components by establishing a temperature-grayscale mapping relationship. Through this precise calculation, not only can the degree of thermal anomaly be quantified, but also a more scientific and accurate basis can be provided for subsequent fault diagnosis and preventive maintenance.

[0196] The technical solution of this application is further described below with reference to specific embodiments:

[0197] According to the technical solution of this application, experiments were conducted to evaluate the thermal hazard identification method of typical transmission line components on a test dataset. The following indicators were mainly used:

[0198] Pixel Accuracy (PA): The PA value obtained in this study is 86.2%, indicating that the vast majority of image pixels are correctly classified.

[0199] Mean Pixel Accuracy (mPA): The mPA value obtained in this study is 84.8%, which reflects the average level of pixel classification in each category.

[0200] Mean Intersection over Union (mIoU): The mIoU value obtained in this study is 78.3%, which proves that the model has high accuracy in segmenting the target area and the background.

[0201] Dice Coefficient: In this study, the Dice value reached 82.5%, further verifying the high overlap between the segmentation results and the true annotations.

[0202] Precision and Recall: In this study, the precision and recall rates were 81.7% and 83.5%, respectively, resulting in a combined F1 score of 92.6%, demonstrating the model's balanced performance in reducing false positives and missed detections.

[0203] Boundary F1 Score: The F1 value in this study is 80.2%, indicating that the model has high accuracy in locating the edges of the segmented objects.

[0204] Accuracy of thermal hazard identification: The accuracy of thermal hazard identification in this study was 86.8%, indicating that the results of curve fitting are highly reasonable.

[0205] Specifically, please refer to Figure 3 The experimental results of the segmentation of typical transmission line components on the test dataset are shown in the figure:

[0206] The first column shows the original image with image region segmentation.

[0207] The second column shows the label mask images.

[0208] The third column shows the segmentation results of the SeA (base) model. This model does not introduce image cues as auxiliary information, and the background area contains a lot of noise, which further weakens the accuracy and clarity of the overall segmentation.

[0209] The fourth column shows the segmentation effect of the SeA(TL) model. This model still does not incorporate additional image clue information, and there is still a certain amount of noise in the background area, which limits the further improvement of the segmentation effect.

[0210] The fifth column shows the segmentation effect of the SeA(MM) model. This model still has certain limitations when relying solely on multimodal data. In some complex environments, the inconsistency of information between modalities may bring additional interference and lead to local segmentation errors.

[0211] The sixth column shows the segmentation effect of the SeA (TL+MM) model. It can be observed that the model is minimally affected by external factors during the segmentation process, and the recognition of typical components of the transmission line is accurate and stable, achieving the optimal segmentation effect. It can be seen that the thermal hazard identification method for typical components of the transmission line provided by this application can automatically capture the complementary characteristics between different modes through multimodal image segmentation: infrared images have advantages in reflecting temperature distribution, while visible light images are more prominent in detail texture and edge information. The organic combination of these two types of information can not only enhance the model's ability to identify the target area, but also effectively suppress background noise and environmental interference, and improve segmentation accuracy.

[0212] See also Figure 4 From the displayed data corresponding to the temperature grayscale values ​​and the result graph of the B-spline curve fitting, it is obvious that the prediction error distribution is relatively uniform, indicating that the relationship between the results of the TG curve fitting and the real data has high repeatability and stability, which further verifies that the application uses the B-spline function to fit the temperature grayscale data to obtain a smooth curve that highly fits the change trend of the real data. This curve can not only reflect the temperature change law corresponding to different grayscale values, but also correct the data deviation caused by factors such as environmental interference or equipment aging to a certain extent.

[0213] The technical solution of this embodiment can achieve efficient identification and monitoring of thermal hazards in transmission lines, thereby discovering and warning of potential equipment failure risks in advance and ensuring the safety and stability of power grid operation. This application uses infrared imaging technology to monitor typical components of transmission lines (such as conductors, insulators, etc.) in real time. By detecting thermal anomaly areas (such as hot spots and local overheating), potential thermal hazards are identified, thereby improving the accuracy and robustness of segmentation of typical components of transmission lines. By adopting a multimodal image segmentation model based on the Transformer architecture and integrating a transfer learning strategy, it can effectively overcome problems such as insufficient single-modal information, noise interference, illumination changes, and image blur, and achieve high-precision segmentation of typical components of transmission lines. At the same time, it improves the accuracy of the temperature grayscale mapping relationship and the reliability of thermal hazard identification. By dynamically fusion of local anomaly factors and isolation forest abnormal data detection methods, interfering data points that may be mixed in during the data extraction process are effectively eliminated, and the cleaned data is accurately fitted using B-spline curves to establish a more accurate temperature grayscale relationship, which enables more accurate calculation of the actual temperature of the component, thereby improving the accuracy of thermal hazard identification and enhancing the scientificity and reliability of diagnosis.

[0214] In the embodiment of the present application, a device for identifying thermal hazards of typical components of a transmission line is provided. Figure 5 , Figure 52 is a structural diagram of a device for identifying thermal hazards of typical components of a transmission line in one embodiment. The device for identifying thermal hazards of typical components of a transmission line includes: a model building module 201, a data processing module 202, an anomaly detection module 203, and a thermal hazard identification module 204.

[0215] The model building module 201 is configured to build a multimodal image segmentation model, and use the multimodal image segmentation model to segment the infrared image containing typical components of the transmission line to obtain segmentation results of the typical components of the transmission line;

[0216] The data processing module 202 is configured to extract the grayscale value of a typical component from the infrared image, and obtain the temperature value corresponding to the grayscale value in combination with preset temperature information to form a temperature-grayscale value data pair;

[0217] An anomaly detection module 203 is configured to perform anomaly detection on the temperature grayscale value data pair, identify and remove abnormal data points, and obtain cleaned temperature grayscale value data;

[0218] The thermal hazard identification module 204 is configured to perform B-spline curve fitting on the cleaned temperature grayscale value data, establish a curve relationship between temperature and grayscale value, and calculate the temperature value of typical components based on the curve relationship to identify thermal hazards of typical components of the transmission line.

[0219] In some embodiments, the model building module 201 is further configured to determine that the multimodal image segmentation model is obtained based on the Transformer architecture, multimodal information and transfer learning strategy; the multimodal image segmentation model includes multi-head self-attention, a feedforward neural network and a classification head.

[0220] In some embodiments, the model building module 201 is further configured to:

[0221] The input infrared image is divided into patches of fixed size, each patch is processed, flattened into a one-dimensional vector, and then converted into a vector of dimension D through linear mapping. The patches are arranged in order to form a sequence, and a learnable classification tag and position encoding are introduced into the sequence to form a final sequence;

[0222] Fusing the visible light image and the infrared image, and inputting the obtained fused image into a feature extraction network to obtain an intermediate feature map F;

[0223] Applying global maximum pooling to the intermediate feature map F and mapping it into prompt information through a feedforward network, and inputting the prompt information into the mask encoding module of infrared image segmentation;

[0224] Encoding the final sequence through a multi-head self-attention and feedforward neural network, and combining the prompt information of the fused image branch to obtain the final encoder output;

[0225] The final encoder output is classified by the classification head to obtain the segmentation result of the infrared image containing typical components of the transmission line.

[0226] In some embodiments, the data processing module 202 is further configured to:

[0227] Determining pixel locations of typical components of the power transmission line based on the segmentation results;

[0228] Extracting the grayscale value at the pixel position from the infrared image, and determining temperature data corresponding to the grayscale value at the pixel position according to a preset mapping relationship between grayscale value and temperature or a calibration curve;

[0229] The grayscale value is combined with the corresponding temperature data to form the temperature grayscale value data pair.

[0230] In some implementations, the anomaly detection module 203 is further configured to:

[0231] Using the isolation forest method to detect the temperature grayscale value data pair, obtain a first anomaly indicator, set a first anomaly score threshold, and screen out a candidate subset containing anomalies that meet the first anomaly score threshold;

[0232] The candidate subset is tested again using the local anomaly factor method to obtain a second anomaly index.

[0233] Assigning a comprehensive anomaly score to each data point in the temperature grayscale value data pair by combining the first anomaly indicator and the second anomaly indicator;

[0234] A second scoring threshold is set according to the comprehensive abnormality score, and abnormal points in the temperature grayscale value data pair are determined and eliminated according to the second scoring threshold to obtain cleaned temperature grayscale value data.

[0235] In some implementations, the anomaly detection module 203 is further configured to:

[0236] Using the isolation forest method to detect the temperature gray value data pair to obtain a first abnormality indicator;

[0237] Detecting the temperature grayscale value data pair using a local anomaly factor method to obtain a second anomaly indicator;

[0238] Assigning a comprehensive anomaly score to each data point in the temperature grayscale value data pair in combination with the first anomaly indicator and the second anomaly indicator;

[0239] A second scoring threshold is set according to the comprehensive abnormality score, and abnormal points in the temperature grayscale value data pair are determined and eliminated according to the second scoring threshold to obtain cleaned temperature grayscale value data.

[0240] In some implementations, the anomaly detection module 203 is further configured to:

[0241] After determining the abnormal points in the temperature grayscale value data pair according to the second scoring threshold, calculating the abnormal proportion;

[0242] If the abnormal ratio does not exceed the preset abnormal ratio threshold, the abnormal points in the temperature gray value data pair are removed;

[0243] If the abnormality ratio exceeds the preset abnormality ratio threshold, the second abnormality scoring threshold is lowered, and the abnormal points in the temperature grayscale value data pair are re-determined and the abnormality ratio is calculated based on the lowered second scoring threshold. When the abnormality ratio does not exceed the preset abnormality ratio threshold, the abnormal points in the temperature grayscale value data pair are eliminated.

[0244] In some embodiments, the thermal hazard identification module 204 is further configured to:

[0245] Determining the number of nodes, wherein the number of nodes refers to the number of uniform nodes or the number of non-uniform nodes;

[0246] Determine the B-spline degree;

[0247] The control points of the temperature grayscale value data are calculated by the least square method, and the curve relationship between the temperature and the grayscale value is determined based on the control points.

[0248] For other details of how the above technical solutions are implemented by each module in the device for identifying thermal hazards of typical components of a transmission line, please refer to the description of the method for identifying thermal hazards of typical components of a transmission line provided above, which will not be repeated here.

[0249] In the embodiment of the present application, a computer device is provided. Figure 6 , Figure 6 3 is a schematic diagram of the structure of a computer device in one embodiment. The device includes a memory 301 and a processor 302. The memory 301 stores a computer program. When the computer program is executed by the processor 302, the processor 302 performs the following steps:

[0250] Constructing a multimodal image segmentation model, and using the multimodal image segmentation model to segment a multimodal image containing typical components of a power transmission line to obtain a segmentation result of the typical components of the power transmission line, wherein the multimodal image includes an infrared image;

[0251] Based on the segmentation result, extracting the grayscale value of the typical component in the infrared image, and obtaining the temperature value corresponding to the grayscale value in combination with the preset temperature information to form a temperature-grayscale value data pair;

[0252] Performing anomaly detection on the temperature grayscale value data pairs, identifying and eliminating abnormal data points, and obtaining cleaned temperature grayscale value data;

[0253] B-spline curve fitting is performed on the temperature grayscale value data after cleaning to establish a curve relationship between temperature and grayscale value, and thermal hazards of typical components of the power transmission line are identified based on the curve relationship.

[0254] Among them, the processor 302 can also be called a CPU (Central Processing Unit), and the processor 302 may be an integrated circuit chip with signal processing capabilities; the processor 302 can also be a general-purpose processor, DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA (Field Programmable Gate Array) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, among which the general-purpose processor can be a microprocessor or the processor 302 can also be any conventional processor, etc.

[0255] In an embodiment of the present application, a computer readable storage medium is provided. Figure 7 , Figure 7 FIG4 is a schematic diagram of the structure of a computer-readable storage medium in one embodiment, wherein the storage medium stores a readable computer program 401; wherein the computer program 401 may be stored in the storage medium in the form of a software product, and includes a number of instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor to perform the following steps:

[0256] Constructing a multimodal image segmentation model, and using the multimodal image segmentation model to segment a multimodal image containing typical components of a power transmission line to obtain a segmentation result of the typical components of the power transmission line, wherein the multimodal image includes an infrared image;

[0257] Based on the segmentation result, extracting the grayscale value of the typical component in the infrared image, and obtaining the temperature value corresponding to the grayscale value in combination with the preset temperature information to form a temperature-grayscale value data pair;

[0258] Performing anomaly detection on the temperature grayscale value data pairs, identifying and eliminating abnormal data points, and obtaining cleaned temperature grayscale value data;

[0259] B-spline curve fitting is performed on the temperature grayscale value data after cleaning to establish a curve relationship between temperature and grayscale value, and thermal hazards of typical components of the power transmission line are identified based on the curve relationship.

[0260] The aforementioned storage media include: U disk, mobile hard disk, magnetic disk or CD, ROM (Read-Only Memory), RAM (Random Access Memory) and other media that can store program code, or terminal devices such as computers, service machines, mobile phones, and tablets.

[0261] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0262] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0263] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for identifying thermal hazards of typical components of a transmission line, characterized in that: include; Constructing a multimodal image segmentation model, and using the multimodal image segmentation model to segment an infrared image containing typical components of a transmission line, to obtain a segmentation result of the typical components of the transmission line; Based on the segmentation result, extracting the grayscale value of the typical component in the infrared image, and obtaining the temperature value corresponding to the grayscale value in combination with the preset temperature information to form a temperature-grayscale value data pair; Performing anomaly detection on the temperature grayscale value data pairs, identifying and eliminating abnormal data points, and obtaining cleaned temperature grayscale value data; B-spline curve fitting is performed on the temperature grayscale value data after cleaning to establish a curve relationship between temperature and grayscale value, and based on the curve relationship, the temperature value of typical components is calculated to identify thermal hazards of typical components of the transmission line.

2. The method for identifying thermal hazards of typical components of a power transmission line according to claim 1, characterized in that: The multimodal image segmentation model is obtained based on the Transformer architecture, multimodal information and transfer learning strategy; the multimodal image segmentation model includes multi-head self-attention, a feedforward neural network and a classification head.

3. The method for identifying thermal hazards of typical components of a power transmission line according to claim 2, characterized in that: The multimodal image segmentation model is constructed, and the infrared image containing typical components of the transmission line is segmented using the multimodal image segmentation model to obtain the segmentation results of the typical components of the transmission line, including: The input infrared image is divided into patches of fixed size, each patch is processed, flattened into a one-dimensional vector, and then converted into a vector of dimension D through linear mapping. The patches are arranged in order to form a sequence, and a learnable classification tag and position encoding are introduced into the sequence to form a final sequence; Fusing the visible light image and the infrared image, and inputting the obtained fused image into a feature extraction network to obtain an intermediate feature map F; Applying global maximum pooling to the intermediate feature map F and mapping it into prompt information through a feedforward network, and inputting the prompt information into the mask encoding module of infrared image segmentation; Encoding the final sequence through a multi-head self-attention and feedforward neural network, and combining the prompt information of the fused image branch to obtain the final encoder output; The final encoder output is classified by the classification head to obtain the segmentation result of the infrared image containing typical components of the transmission line.

4. The method for identifying thermal hazards of typical components of a power transmission line according to claim 3, characterized in that: The preset temperature information includes a mapping relationship or a calibration curve between a preset grayscale value and temperature; The grayscale value is extracted from the original infrared image based on the segmentation result, and the corresponding temperature value is obtained in combination with the preset temperature information to form a temperature grayscale value data pair, including: Determining pixel locations of typical components of the power transmission line based on the segmentation results; Extracting the grayscale value at the pixel position from the infrared image, and determining temperature data corresponding to the grayscale value at the pixel position according to a preset mapping relationship between grayscale value and temperature or a calibration curve; The grayscale value is combined with the corresponding temperature data to form the temperature grayscale value data pair.

5. The method for identifying thermal hazards of typical components of a power transmission line according to claim 4, characterized in that: The performing abnormality detection on the temperature grayscale value data, identifying and eliminating abnormal data points, and obtaining cleaned temperature grayscale value data includes: Using the isolation forest method to detect the temperature grayscale value data pair, obtain a first anomaly indicator, set a first anomaly score threshold, and screen out a candidate subset containing anomalies that meet the first anomaly score threshold; The candidate subset is tested again using the local anomaly factor method to obtain a second anomaly index. Assigning a comprehensive anomaly score to each data point in the temperature grayscale value data pair by combining the first anomaly indicator and the second anomaly indicator; A second scoring threshold is set according to the comprehensive abnormality score, and abnormal points in the temperature grayscale value data pair are determined and eliminated according to the second scoring threshold to obtain cleaned temperature grayscale value data.

6. The method for identifying thermal hazards of typical components of a power transmission line according to claim 5, characterized in that: The determining and eliminating abnormal points in the temperature grayscale value data pair according to the second scoring threshold comprises: After determining the abnormal points in the temperature grayscale value data pair according to the second scoring threshold, calculating the abnormal proportion; If the abnormal ratio does not exceed the preset abnormal ratio threshold, the abnormal points in the temperature gray value data pair are removed; If the abnormality ratio exceeds the preset abnormality ratio threshold, the second abnormality scoring threshold is lowered, and the abnormal points in the temperature grayscale value data pair are re-determined and the abnormality ratio is calculated based on the lowered second scoring threshold. When the abnormality ratio does not exceed the preset abnormality ratio threshold, the abnormal points in the temperature grayscale value data pair are eliminated.

7. The method for identifying thermal hazards of typical components of a power transmission line according to claim 6, characterized in that: The performing B-spline curve fitting on the temperature grayscale value data after cleaning includes: Determining the number of nodes, wherein the number of nodes refers to the number of uniform nodes or the number of non-uniform nodes; Determine the B-spline degree; The control points of the temperature grayscale value data are calculated by the least square method, and the curve relationship between the temperature and the grayscale value is determined based on the control points.

8. A device for identifying thermal hazards of typical components of a power transmission line, characterized in that: The device comprises: a model building module, configured to build a multimodal image segmentation model, and use the multimodal image segmentation model to segment infrared images containing typical components of a transmission line, thereby obtaining segmentation results of the typical components of the transmission line; a data processing module, configured to extract the grayscale value of a typical component from the infrared image, and obtain the temperature value corresponding to the grayscale value in combination with preset temperature information, thereby forming a temperature-grayscale value data pair; An anomaly detection module is used to perform anomaly detection on the temperature gray value data pair, identify and eliminate abnormal data points, and obtain cleaned temperature gray value data; The thermal hazard identification module is used to perform B-spline curve fitting on the temperature grayscale value data after cleaning, establish a curve relationship between temperature and grayscale value, and calculate the temperature value of typical components based on the curve relationship to identify thermal hazards of typical components of the transmission line.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images