An ultra-high voltage line forest fire prediction method based on multi-source data fusion
By integrating multi-source data and using multi-network collaborative modeling, the problems of incomplete feature extraction and insufficient generalization ability of single data sources and single network models in monitoring wildfires on ultra-high voltage lines have been solved, achieving more accurate and reliable wildfire prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGXI UNIV
- Filing Date
- 2026-03-10
- Publication Date
- 2026-06-05
AI Technical Summary
Existing methods for monitoring wildfires along ultra-high voltage power lines rely on single image data, resulting in incomplete feature extraction. Single network models have limited feature extraction capabilities and weak generalization ability, making it difficult to accurately predict wildfires in complex environments.
By employing multi-source image data fusion and multi-network collaborative modeling, and constructing multiple sets of parallel training systems, the network parameters with the best overall performance were selected. Combined with ResNet18, VGG19, InceptionV3, DarkNet19 and deep network models, wildfire characteristics were comprehensively captured and the model's adaptability was improved.
It improves the accuracy and reliability of wildfire prediction, reduces misjudgments and omissions caused by environmental interference, and enhances the model's predictive ability in complex scenarios.
Smart Images
Figure CN122156977A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of power system safety monitoring, image recognition, and deep learning applications. It involves a multi-source data fusion method and is applicable to wildfire safety monitoring and early warning of ultra-high voltage power lines. Background Technology
[0002] In the current field of ultra-high voltage power line wildfire monitoring, existing prediction methods mostly rely on single-type image data for analysis, using only satellite remote sensing images or images taken by a single camera as model input. This results in incomplete extraction of wildfire-related features and makes it difficult to capture the detailed features and dynamic changes of wildfires from different data sources.
[0003] Furthermore, existing prediction methods often employ a single network model to complete the classification task, and the feature extraction capabilities of a single network have inherent limitations: methods relying on residual networks can alleviate the gradient vanishing problem, but their adaptability to multi-scale features is insufficient; methods based on lightweight networks, while computationally efficient, have limited feature extraction depth and struggle to handle complex scenarios such as vegetation cover, terrain shadows, and atmospheric interference; methods using multi-branch convolutional networks can capture some multi-scale features, but they do not fully incorporate cross-time data differences, resulting in low sensitivity to feature changes during wildfire development and a high risk of misclassification or underclassification. Simultaneously, existing models do not fully explore the optimization space of multiple sets of training parameters during training, leading to weak model generalization capabilities. When facing wildfire scenarios involving ultra-high voltage power lines in different regions and under different climatic conditions, they lack adaptability and struggle to consistently output reliable prediction results. Summary of the Invention
[0004] This invention proposes a method for predicting wildfires along ultra-high voltage power lines based on multi-source data fusion. It combines multi-source image data fusion with multi-network collaborative modeling for the classification and prediction of wildfires along ultra-high voltage power lines. This method comprehensively captures both static and dynamic features of wildfires, improves the completeness of feature extraction in complex scenarios, and enhances the model's generalization ability. It can improve the accuracy and reliability of wildfire prediction along ultra-high voltage power lines and reduce misjudgments and missed judgments caused by environmental interference. The steps in its application are as follows: Step (1): Construct multiple parallel training systems. All parallel training systems are aimed at predicting wildfires along ultra-high voltage power lines. They use completely identical training data, sample sets, and network model architectures, and only set different model training parameters to carry out independent training. After training, the model performance is comprehensively evaluated by accuracy, precision, recall, and F1 score. The five network models with the best overall performance are selected, and their trained network parameters are extracted and transferred to the actual ultra-high voltage power line wildfire prediction system. The simultaneous operation of multiple parallel training systems can quickly find the parameters required by multiple network models in the actual ultra-high voltage power line wildfire prediction system. The simultaneous operation of multiple parallel training systems belongs to the artificial society-computational experiment-parallel execution framework. Step (2): Collect four types of image pixel data, specifically: Original satellite image: The pixel matrix of the original satellite image of the ultra-high voltage line monitoring area at the current moment, denoted as... Figure 1 Image size is N 1× N 2×3, for example, you can choose N 1 = 224 N 2 = 224, the pixel value range is [0, 255], and the pixel matrix dimension corresponds one-to-one with the geographical range of the monitoring area; Original camera image: The pixel matrix of the original image captured by the ultra-high voltage line monitoring camera at the current moment, denoted as... Figure 2 Image size is N 1× N 2×3, for example, you can choose N 1 = 224 N 2=224, pixel value range is [0,255], shooting angle covers key monitoring sections of ultra-high voltage lines; Satellite difference map: The difference pixel matrix is obtained by subtracting the pixel matrix of the original satellite image at the current time from the pixel matrix of the original satellite image at the previous time, pixel by pixel. The pixel value range is [-255, 255]. This difference pixel matrix is then normalized to the interval [0, 255] and denoted as . Figure 3 Matrix dimension and satellite original Figure 1 This is used to reflect pixel changes in the monitored area over time. Camera difference map: The difference pixel matrix is obtained by subtracting the pixel matrix of the original camera image at the current moment from the pixel matrix of the original camera image at the previous moment pixel by pixel. The pixel value range is [-255, 255]. This difference matrix is then normalized to the interval [0, 255] and denoted as . Figure 4 Matrix dimension and camera original Figure 1 This is used to capture the dynamic changes in the on-site environment; Step (3): Figure 1 , Figure 2 , Figure 3 , Figure 4 The inputs are fed into network model 1, network model 2, network model 3, and network model 4 respectively. For example, ResNet18, VGG19, InceptionV3, DarkNet19, etc. can be selected respectively. Any convolutional network that can process images can be used as the model for the four networks. The four networks extract features and output classification probabilities for the input pixel matrix respectively. Step (4): Combine the three-class classification probability values output by the four networks in step (3) into a one-dimensional feature vector in order. The feature vector has a dimension of 12, where the first 3 dimensions are the output probability of network model 1, the middle 3 dimensions are the output probability of network model 2, and the last 6 dimensions are the output probabilities of network model 3 and network model 4, respectively. Step (5): Input the 12-dimensional feature vector obtained in step (4) into the deep network model 5. For example, the deep network model 5 can be composed of a feature input layer, a fully connected layer, a softmax layer and a classification layer. The feature input layer of the fusion network receives the 12-dimensional feature vector and performs standardization processing. Networks with feature input and classification output can be used as the network model of network 5. The ReLU activation function is used for nonlinear feature fusion. The softmax layer maps the fused features to the probability value of the nc-class ultra-high voltage line wildfire classification. For example, when nc=3, it can be divided into "no fire", "smoke" and "smoke". For example, when nc=5, it can be divided into no fire, light fog / dust, smoke, small fire spot and smoke. The classification layer selects the category corresponding to the maximum probability as the final classification result.
[0005] The present invention has the following advantages and effects compared with the prior art: (1) Existing methods rely on single image data and feature extraction is not comprehensive. This invention collects four types of pixel data, including original satellite images, original camera images, and two types of cross-time difference images, and integrates multi-source image information to comprehensively capture the static and dynamic features of wildfires, thus solving the problem of insufficient features from a single data source.
[0006] (2) Existing methods use a single network model, which has limited feature extraction capabilities. This invention combines the advantages of four networks, ResNet18, VGG19, InceptionV3, and DarkNet19, to extract features from different dimensions of multi-source data. Then, the probability output is integrated through a fusion network to improve the feature capture capability and classification reliability in complex scenarios.
[0007] (3) Existing methods have weak generalization ability. This invention constructs multiple parallel training systems, selects the five network parameters with the best accuracy for the actual system, optimizes the model training effect, enhances the model's adaptability to different scenarios, and improves the stability of prediction results.
[0008] (4) Existing methods have insufficient prediction accuracy in complex environments. This invention effectively resists the influence of environmental factors such as vegetation cover, terrain shadows, and atmospheric interference by multi-source data fusion and multi-network collaborative prediction, improves the accuracy of wildfire prediction for ultra-high voltage lines, and provides reliable support for early warning of wildfires. Attached Figure Description
[0009] Figure 1This is a parallel system framework diagram of the method of the present invention.
[0010] Figure 2 This is a flowchart of the multi-source data network fusion method of the present invention.
[0011] Figure 3 This is a diagram of the ResNet18 layered architecture of the method of this invention.
[0012] Figure 4 This is a VGG19 layered architecture diagram of the method of this invention.
[0013] Figure 5 This is a diagram of the InceptionV3 layered architecture of the method of this invention.
[0014] Figure 6 This is a diagram of the DarkNet19 layered architecture of the method of this invention.
[0015] Figure 7 This is a diagram of the layered architecture of the deep network model of the method of this invention. Detailed Implementation
[0016] This invention proposes a method for predicting wildfires along ultra-high voltage power lines based on multi-source data fusion, which is described in detail below with reference to the accompanying drawings: Figure 1 This is a parallel system framework diagram of the method of this invention. First, multiple parallel training systems are constructed with the goal of predicting wildfires along ultra-high voltage power lines. All parallel training systems are configured with completely identical training data, sample sets, and network model architectures, with only differentiated model training parameters for independent training. Then, the performance of all trained parallel systems is comprehensively evaluated using accuracy, precision, recall, and F1 score, and the five network models with the best overall performance are selected. Finally, the post-training parameters of these five optimal network models are extracted and transferred to the actual ultra-high voltage power line wildfire prediction system as the initial operating parameters for that system.
[0017] Figure 2 This is a flowchart of the multi-source data network fusion method of the present invention. First, four types of image pixel data from the monitoring area of the ultra-high voltage line are collected: original satellite image, original camera image, satellite differential image, and camera differential image. Both the satellite differential image and the camera differential image are normalized to the pixel value range of [0, 255]. Then, the four types of data are input into network models 1, 2, 3, and 4 according to their corresponding relationships. Each of the four networks extracts features from the input pixel matrix and outputs three classification probability values. Finally, the probability values output by the four networks are combined sequentially into a 12-dimensional feature vector, which is input into a deep network model 5. The classification layer selects the category corresponding to the maximum probability as the final prediction result for wildfires along the ultra-high voltage line.
[0018] Figure 3 This is a diagram of the ResNet18 layered architecture of the method of this invention. First, the input... N 1× N A 2×3 image is passed through a 7×7 convolutional layer with a stride of 2 and padding of 3, resulting in 64 output channels, extracting low-level features such as image edges and corners. Next, multiple residual blocks are passed, each containing two convolutional layers and a residual connection. Within each residual block, the first convolutional layer uses a 3×3 kernel with a stride of 1 and padding of 1, and the same number of output channels as the input. The second convolutional layer also uses a 3×3 kernel with a stride of 1 and padding of 1, but the number of output channels is twice the number of input channels. Finally, after passing through multiple residual blocks, the feature map size becomes 7×7 with 512 channels. The feature map is flattened and fed into a fully connected layer, outputting the final classification result.
[0019] Figure 4 This is a VGG19 layered architecture diagram of the method of this invention. First, the input... N 1× N The 2×3 image is sequentially processed through 5 sets of stacked convolutional units to extract basic features. Each set of units consists of multiple 3×3 convolutional layers connected in series. At the end of each set of convolutional units, a 2×2 max pooling layer is used to downsample the feature map, gradually reducing the feature map size and increasing the number of channels. Finally, after processing by all convolutional and pooling units, the feature map size becomes 7×7 with 512 channels. The feature map is then flattened and input into a fully connected layer to output the final classification result.
[0020] Figure 5 This is a diagram of the InceptionV3 layered architecture of the method of this invention. First, the input... N 1× N The 2×3 image is processed through three 3×3 convolutional layers, one 3×3 max-pooling layer, one 1×1 convolutional layer, one 3×3 convolutional layer, and one 3×3 max-pooling layer to extract basic shallow features. Next, multiple improved Inception modules are used, each containing four parallel convolutional branches. Each branch employs 1×1, 3×3, and 5×5 convolutional kernels, as well as a combination of "3×3 pooling + 1×1 convolution," to achieve channel dimensionality reduction and feature fusion. Finally, after passing through multiple Inception modules, the feature map size becomes 7×7 with 2048 channels. The feature map is flattened and input into a fully connected layer, outputting the final classification result.
[0021] Figure 6 This is a diagram of the DarkNet19 layered architecture of the method of this invention. First, the input... N 1× NThe 2×3 image is sequentially processed through five sets of convolutional pooling units to complete basic feature extraction and downsampling. Each unit consists of a combination of a 1×1 convolutional layer and a 3×3 convolutional layer, achieving channel dimensionality reduction and cross-channel feature fusion. At the end of each unit, a 2×2 max-pooling layer completes feature map downsampling. Next, four sets of unpooled 1×1+3×3 convolutional layers are used to further extract deeper image features. Finally, after processing by all convolutional layers, the feature map size becomes 7×7 with 1024 channels. The feature map is flattened and input into a fully connected layer, outputting the final classification result.
[0022] Figure 7 This is a diagram of the 5-layer architecture of the deep network model of the method of this invention. The obtained 12-dimensional feature vector is input and passed sequentially through the feature input layer, the fully connected layer, and the Softmax layer to output the final classification result.
[0023] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for predicting wildfires along ultra-high voltage power lines based on multi-source data fusion, comprising the following steps: Step (1): A plurality of parallel training systems are constructed, all of which are aimed at ultra-high voltage line forest fire prediction, use completely consistent training data, sample sets and network model architectures, and only set different model training parameters to carry out independent training; After training, the model performance is comprehensively evaluated by accuracy, precision, recall and F1 score. The five network models with the best overall performance are selected, and their trained network parameters are extracted and transferred to the actual ultra-high voltage line wildfire prediction system. Step (2): Collect four types of image pixel data, specifically: Original satellite image: The pixel matrix of the original satellite image of the ultra-high voltage line monitoring area at the current moment, denoted as Figure 1, with an image size of [missing information]. N 1× N 2×3, with pixel values ranging from [0,255], and the pixel matrix dimensions correspond one-to-one with the geographical range of the monitoring area; Original camera image: The pixel matrix of the original image captured by the ultra-high voltage line monitoring camera at the current moment, denoted as Figure 2, with an image size of [missing information]. N 1× N 2×3, pixel value range [0,255], shooting angle covers key monitoring sections of ultra-high voltage lines; Satellite difference map: The difference pixel matrix is obtained by subtracting the pixel matrix of the original satellite image at the current time from the pixel matrix of the original satellite image at the previous time pixel by pixel. The pixel value range is [-255, 255]. It is then normalized to the interval [0, 255], as shown in Figure 3. The matrix dimension is the same as that of the original satellite image. It is used to reflect the pixel change information of the monitoring area across time. Camera difference map: The difference pixel matrix is obtained by subtracting the original pixel matrix of the camera at the current moment from the original pixel matrix of the camera at the previous moment pixel by pixel. The pixel value range is [-255, 255]. It is then normalized to the interval [0, 255], as shown in Figure 4. The matrix dimension is the same as that of the original camera image, and it is used to capture the dynamic change features of the scene environment. Step (3): Input Figure 1, Figure 2, Figure 3 and Figure 4 into Network Model 1, Network Model 2, Network Model 3 and Network Model 4 respectively. The four networks extract features and output classification probabilities for the input pixel matrix respectively. Step (4): Combine the three-class classification probability values output by the four networks in step (3) into a one-dimensional feature vector in order. The feature vector has a dimension of 12, where the first 3 dimensions are the output probability of network model 1, the middle 3 dimensions are the output probability of network model 2, and the last 6 dimensions are the output probabilities of network model 3 and network model 4, respectively. Step (5): Input the 12-dimensional feature vector obtained in step (4) into the deep network model 5, and use the ReLU activation function to perform nonlinear feature fusion. The fused feature mapping is the probability value of the wildfire classification of nc-class ultra-high voltage lines. The classification layer selects the category corresponding to the maximum probability as the final classification result.