A method for identifying wildfires in power transmission corridors based on deep learning

Abstract

The present invention belongs to the technical field of power transmission corridor wildfire identification, specifically a power transmission corridor wildfire identification method based on deep learning, including the following specific steps: constructing a power transmission corridor wildfire image data set; performing feature enhancement on the power transmission corridor wildfire images in the power transmission corridor wildfire image data set; constructing a power transmission corridor wildfire identification model based on deep learning, inputting the power transmission corridor wildfire images into the model, and the model automatically identifies the wildfires in the power transmission corridor images; dividing the power transmission corridor wildfire image data set, training and verifying the power transmission corridor wildfire identification model; applying the power transmission corridor wildfire identification model, the present invention extracts multi-level feature map information so that the model can more accurately extract wildfire features, thereby effectively improving the accuracy of wildfire identification in a complex power transmission corridor environment, and through continuous learning and training, the accuracy and efficiency of wildfire identification can be further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wildfire identification in power transmission corridors, and specifically to a method for wildfire identification in power transmission corridors based on deep learning. Background Art

[0002] The transmission corridor, also known as the transmission line corridor or high-voltage line corridor, refers to a strip area under the line with a specified width extending to both sides along the side conductors of the high-voltage overhead power line. With the influence of global climate change and human activities, the frequency and intensity of wildfires in the transmission corridor are increasing year by year, posing a serious threat to the safe and stable operation of the power system. Once the transmission corridor is affected by a wildfire, it may lead to serious consequences such as transmission line tripping and tower burning, and may even cause large-scale power outages, causing huge losses to social production and people's lives. Therefore, designing an efficient and accurate method for identifying wildfires in transmission corridors is of great significance to ensure the safe and stable operation of the power system.

[0003] At present, traditional image processing technology is not ideal for wildfire identification in complex power transmission corridor environments and is difficult to meet the needs of practical applications. The specific defects are as follows:

[0004] 1. The existing technology uses traditional machine learning algorithms, which have low recognition accuracy and are prone to missed detection and false detection;

[0005] 2. Existing technologies have limited ability to accurately identify, classify, and locate complex wildfire scenarios, such as outdoor transmission corridor scenarios.

[0006] Therefore, a method for identifying wildfires in transmission corridors based on deep learning is invented. The powerful feature extraction and classification capabilities of deep learning technology can be used to automatically identify wildfires in transmission corridors. By training the deep learning model, it can automatically learn and identify the characteristics of wildfires, thereby achieving rapid and accurate identification of wildfires and ensuring the safety and stability of the power system. Summary of the invention

[0007] In view of the above-mentioned and / or existing problems in a method for identifying wildfires in power transmission corridors based on deep learning, the present invention is proposed.

[0008] Therefore, the purpose of the present invention is to provide a method for identifying wildfires in power transmission corridors based on deep learning, which can solve the above-mentioned existing problems.

[0009] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0010] A method for identifying wildfires in power transmission corridors based on deep learning includes the following specific steps:

[0011] S1: Construct a dataset of wildfire images in power transmission corridors;

[0012] S2: Feature enhancement of transmission corridor wildfire images in the transmission corridor wildfire image dataset;

[0013] S3: A transmission corridor wildfire recognition model based on deep learning was built. The wildfire images of transmission corridors were input into the model, and the model automatically identified the wildfires in the transmission corridor images.

[0014] S4: Divide the transmission corridor wildfire image dataset, train and verify the transmission corridor wildfire recognition model;

[0015] S5: Apply the transmission corridor wildfire identification model.

[0016] As a preferred solution of the method for identifying wildfires in power transmission corridors based on deep learning described in the present invention, the specific process of S1 is as follows:

[0017] S11: Collect transmission corridor wildfire images or use the public transmission corridor wildfire image dataset to build a transmission corridor wildfire image dataset;

[0018] S12: The collected transmission corridor wildfire image dataset is manually labeled, and the LabelImg image labeling tool is used to select the wildfire location in the transmission corridor wildfire image and label its category as wildfire.

[0019] As a preferred solution of the method for identifying wildfires in power transmission corridors based on deep learning described in the present invention, the specific process of S2 is as follows:

[0020] S21: Processing the transmission corridor wildfire image in the transmission corridor wildfire image dataset into an RGB image;

[0021] S22: Enhance the wildfire features in the transmission corridor wildfire image dataset by increasing the R element in the image and reducing the values ​​of the G and B elements;

[0022] The processing calculation formula is as follows:

[0023] R_adj(x,y)=max(0,min(255,R_orig(x,y)×α))

[0024] G_adj(x,y)=max(0,min(255,G_orig(x,y)×β))

[0025] B_adj(x,y)=max(0,min(255,B_orig(x,y)×β))

[0026] Among them, R_orig(x,y), G_orig(x,y), and B_orig(x,y) represent the R, G, and B element values ​​of the original image at the (x,y) position, ranging from 0 to 255. R_adj(x,y), G_adj(x,y), and B_adj(x,y) represent the R, G, and B element values ​​of the adjusted image at the (x,y) position. α is used to increase the factor of the R element, which is a number greater than 1. β is used to reduce the factors of the G and B elements, which is a number between 0 and 1. Increasing the value of the R element can significantly enhance the red feature of the flame in the image, making it more prominent and easier to identify. At the same time, reducing the values ​​of the G and B elements can help reduce the interference of the background and other non-flame areas, further highlight the flame area, and improve the contrast of the overall image. Then the image size is uniformly expanded or compressed to 512×512×3 to ensure the consistency of the size of the subsequent input images.

[0027] As a preferred solution of the method for identifying wildfires in power transmission corridors based on deep learning described in the present invention, the specific process of S3 is as follows:

[0028] S31: Design a DFM dynamic perception feature extraction module, which expands the detail feature receptive field of the transmission corridor wildfire image through dual-branch dynamic convolution operation, so as to better capture the transmission corridor wildfire information in the image and enhance the model's recognition ability of transmission corridor wildfire;

[0029] S32: Design a DF-SPPF pyramid pooling structure so that the DF-SPPF pyramid pooling structure can extract the features of the power transmission corridor wildfire images using parallel dilated convolution branches, and fuse the features of the power transmission corridor wildfire images of different scales through pooling operations, thereby improving the model's wildfire recognition ability for the power transmission corridor wildfire images of different scales;

[0030] S33: Design an AD-Head hole detection head so that the AD-Head hole detection head can increase the receptive field of each convolution kernel while maintaining the size of the feature map, which helps to extract more context information from the input feature map and capture target features of different scales, so that the AD-Head hole detection head can extract rich feature information, thereby improving detection accuracy and efficiency;

[0031] S34: A transmission corridor wildfire recognition model is proposed using the DFM dynamic perception feature extraction module designed in S31, the DF-SPPF pyramid pooling structure designed in S32, and the AD-Head void detection head designed in S33. By inputting the transmission corridor wildfire image, the wildfire in the image is automatically identified and located.

[0032] As a preferred solution of the method for identifying wildfires in power transmission corridors based on deep learning described in the present invention, the specific process of S31 is as follows:

[0033] S311: Input the transmission corridor wildfire characteristic map T with size H×W×C 1 The convolution kernel number is C, the size is 3×3, and the padding value is 2. The transmission corridor wildfire feature map T with the size of H×W×C is obtained. 2 ;

[0034] S312: Transmission corridor wildfire characteristic map T with size H×W×C 2 , split along the channel to obtain the transmission corridor wildfire characteristic map T with a size of H×W×C / 2 3 and T 4 ;

[0035] S313: Transmission corridor wildfire characteristic map T 3 After the average pooling operation, the image is input into the dynamic convolution, followed by the BN batch normalization operation and the Swish activation function for nonlinear transformation. The result is then input into the GAM to obtain the transmission corridor wildfire feature map T with a size of H×W×C / 2. 5 ;

[0036] T 5 =GAM(Swish(BN(DConv(AVGPooling(T 3 )))))

[0037] AVGPooling stands for average pooling operation, DConv stands for dynamic convolution, BN stands for batch normalization, and GAM stands for global attention mechanism.

[0038] S314: Transmission corridor wildfire characteristic map T 4 After the maximum pooling operation, it is input into the dynamic convolution, followed by the BN batch normalization operation and the Swish activation function for nonlinear transformation, and then input into the GAM to obtain the transmission corridor wildfire feature map T with a size of H×W×C / 2. 6 ;

[0039] T 6 =GAM(Swish(BN(DConv(MAXPooling(T 4 )))))

[0040] Among them, MAXPooling represents the maximum pooling operation, DConv represents dynamic convolution, BN represents batch normalization, and GAM represents the global attention mechanism;

[0041] S315: Transmission corridor wildfire characteristic map T with size H×W×C / 25 and T 6 Superimpose along the channel to obtain the transmission corridor wildfire characteristic map T with a size of H×W×C 7 , T 7 Input into the CBS module with C convolution kernels, size 3×3, and padding value 2, and obtain and output the transmission corridor wildfire feature map T with size H×W×C. 8 .

[0042] As a preferred solution of the method for identifying wildfires in power transmission corridors based on deep learning described in the present invention, the specific process of S32 is as follows:

[0043] S321: Input transmission corridor wildfire feature map A with size H×W×C 1 In the first branch, A 1 After the calculation is performed with a dilated convolution with C convolution kernels, a size of 3×3, a padding value of 2, and an expansion rate of 2, the results are sequentially input into the BN batch normalization layer and the ReLu activation function to obtain a transmission corridor wildfire feature map A with a size of H×W×C. 2 , in the second branch A 1 After the calculation of the dilated convolution with C convolution kernels, size 3×3, padding value 2, and expansion rate 4, the results are input into the BN batch normalization layer and ReLu activation function in turn to obtain the transmission corridor wildfire feature map A with size H×W×C. 3 , in the third branch A 1 After the calculation of the dilated convolution with C convolution kernels, size 3×3, padding value 2, and expansion rate 6, the results are input into the BN batch normalization layer and ReLu activation function in turn to obtain the transmission corridor wildfire feature map A with size H×W×C. 4 ;

[0044] A 2 =ReLU(BN(DLConv r＝2 (A 1 )))

[0045] A 3 =ReLU(BN(DLConv r＝4 (A 1 )))

[0046] A 4 =ReLU(BN(DLConv r＝6 (A 1 )))

[0047] Among them, DLConv r＝2 , DLConv r＝4 , DLConvr＝6 They represent the calculation of dilated convolution with expansion rates of 2, 4, and 6 respectively, BN represents batch normalization, and ReLU is the activation function;

[0048] S322: Transmission corridor wildfire feature map A with dimensions H×W×C 2 , A 3 , A 4 The transmission corridor wildfire characteristic map A with the size of H×W×3C is obtained by superimposing along the channel. 5 , then A 5 Input into the CBS module with C convolution kernels, size 3×3, and padding value 2, and obtain the transmission corridor wildfire feature map A with size H×W×C 6 ;

[0049] A 5 =Concat(A 2 , A 3 , A 4

[0050] Concat is a superposition operation along the channel;

[0051] S323: A 6 Input to three consecutive maximum pooling layers to obtain the transmission corridor wildfire feature map A with a size of H×W×C. 7 , A 8 , A 9 , the transmission corridor wildfire feature map A with the size of H×W×C 6 , A 7 , A 8 , A 9 The transmission corridor wildfire characteristic map A with the size of H×W×4C is obtained by superimposing along the channel. 10 , then A 10 Input into the CBS module with C convolution kernels, size 3×3, and padding value 2, and obtain and output the transmission corridor wildfire feature map A with size H×W×C 11 ;

[0052] A 7 =MAXPooling(A 6 )

[0053] A 8 =MAXPooling(A 7 )

[0054] A 9 =MAXPooling(A 8 )

[0055] A 10 =Concat(A 6, A 7 , A 8、 A 9 ).

[0056] As a preferred solution of the method for identifying wildfires in power transmission corridors based on deep learning described in the present invention, the specific process of S33 is as follows:

[0057] S331: Input the transmission corridor wildfire feature map L with size H×W×C 1 To the CBS module with C convolution kernels, size 3×3, and padding value 2, we get the transmission corridor wildfire feature map L with size H×W×C 2 ;

[0058] S332: L 2 Input into three branches. In the first branch, L 2 First, the transmission corridor wildfire feature map L with a size of H×W×C is obtained through a CBS module with a convolution kernel number of C, a size of 3×3, and a padding value of 2. 3 , then L 3 Input into a two-dimensional convolution layer with C convolution kernels and a size of 1×1 to obtain the H×W×C transmission corridor wildfire feature map L 4 , L 4 The target category of the characteristic image of wildfire in the power transmission corridor is obtained through calculation;

[0059] S333: In the second branch, L 2 The input is a dilated convolution with a kernel size of C, a size of 3×3, and an expansion rate of 2, and then BN batch normalization and ReLu activation function operations are performed to obtain a transmission corridor wildfire feature map L with a size of H×W×C. 5 , in the third branch, L 2 The input is a dilated convolution with a kernel size of C, a size of 3×3, and an expansion rate of 4, and then BN batch normalization and ReLu activation function operations are performed to obtain a transmission corridor wildfire feature map L with a size of H×W×C. 6 , L 5 and L 6 Add the features to get the transmission corridor wildfire feature map L with the size of H×W×C 7 , then L 7 Enter into two branches, in the upper branch L 7 The input is sent to a two-dimensional convolutional layer with 4 convolution kernels and a size of 1×1 to obtain the transmission corridor wildfire feature map L with a size of H×W×4. 8 , L 8 After calculating, the wildfire target regression frame of the wildfire feature image of the transmission corridor is obtained. In the lower branch, L 7Input into a two-dimensional convolution layer with 1 convolution kernel and 1×1 size to obtain the transmission corridor wildfire feature map L with a size of H×W×1 9 , L 9 The wildfire target confidence of the wildfire feature image in the transmission corridor is obtained through calculation.

[0060] As a preferred solution of the method for identifying wildfires in power transmission corridors based on deep learning described in the present invention, the specific process of S34 is as follows:

[0061] S341: Input a transmission corridor wildfire image F with a size of 512×512×3 1 The convolution kernel number is 64, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 256×256×64 is obtained. 2 , and then F 2 Input into the CBS module with 128 convolution kernels, 3×3 size, 2 steps, and 1 padding value, and obtain the transmission corridor wildfire feature map F with a size of 128×128×128 3 ;

[0062] S342: F 3 The input convolution kernel number is 256, the size is 3×3, the step length is 2, and the padding value is 1. Then it is input into the DFM dynamic perception feature extraction module for feature extraction to obtain the transmission corridor wildfire feature map F with a size of 64×64×256. 4 , F 4 The input convolution kernel number is 512, the size is 3×3, the step size is 2, and the padding value is 1. Then it is input into the DFM dynamic perception feature extraction module for feature extraction to obtain the transmission corridor wildfire feature map F with a size of 32×32×512. 5 , F 5 Input the CBS module with 512 convolution kernels, 3×3 kernel size, 2 stride, and 1 padding value, and then input DFM to obtain the transmission corridor wildfire feature map F with a size of 16×16×512. 6 , F 6 Input into the DF-SPPF pyramid pooling structure to obtain the transmission corridor wildfire feature map F with a size of 16×16×512 7 ;

[0063] S343: F 7 The upsampling operation is performed to obtain the transmission corridor wildfire feature map F with a size of 32×32×512 8 , F 5 and F 8After superimposing along the channel, the map is input into the DFM dynamic perception feature extraction module to obtain the transmission corridor wildfire feature map F with a size of 32×32×1024. 9 ;

[0064] S344: F 9 The upsampling operation is performed to obtain the transmission corridor wildfire feature map F with a size of 64×64×512 10 , F 4 and F 10 After superimposing along the channel, the map is input into the DFM dynamic perception feature extraction module to obtain the transmission corridor wildfire feature map F with a size of 64×64×768. 11 ;

[0065] S345: F 11 The number of input convolution kernels is 256, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 32×32×256 is obtained. 12 , F 9 and F 12 After superimposing along the channel, the map is input into the DFM dynamic perception feature extraction module to obtain the transmission corridor wildfire feature map F with a size of 64×64×1280. 13 ;

[0066] S346: Transmission Corridor Wildfire Characteristic Map F 13 The number of input convolution kernels is 512, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 32×32×512 is obtained. 14 , F 8 and F 14 After superimposing along the channel, the map is input into the DFM dynamic perception feature extraction module to obtain the transmission corridor wildfire feature map F with a size of 32×32×1024. 15 ;

[0067] S347: Transmission Corridor Wildfire Characteristic Map F 11 The number of input convolution kernels is 256, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 32×32×256 is obtained. 16 , the transmission corridor wildfire characteristic map F 13 The number of input convolution kernels is 512, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 16×16×512 is obtained. 17 , the transmission corridor wildfire characteristic map F 15The number of input convolution kernels is 1024, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 8×8×1024 is obtained. 18 ;

[0068] S348: Transmission corridor wildfire feature map F 16 、F 17 、F 18 The image is input into the AD-Head hole detection head to obtain and output the wildfire recognition result, which includes the target category, regression box and confidence of the wildfire in the image.

[0069] As a preferred solution of the method for identifying wildfires in power transmission corridors based on deep learning described in the present invention, the specific process of S4 is as follows:

[0070] S41: Divide the transmission corridor wildfire image dataset constructed in S1 into a training set and a validation set at a ratio of 4:1, and train the transmission corridor wildfire recognition model constructed in S3;

[0071] S42: Initialize all neural network parameters and set model-related training parameters, divide the training set and validation set data into multiple batches, input a batch of training set data into the algorithm for training each time, and obtain the training loss value of the batch;

[0072] S43: Input the validation set into the model in batches to obtain the corresponding batch loss value, automatically learn and adjust the parameters according to the batch loss value each time, until the batch loss value tends to converge, and the training of the transmission corridor wildfire recognition model is completed.

[0073] As a preferred solution of the method for identifying wildfires in power transmission corridors based on deep learning described in the present invention, the specific process of S5 is as follows:

[0074] S51: After the training of the transmission corridor wildfire identification model is completed, the model is integrated into the existing transmission corridor monitoring system to ensure that the model can receive image or video data of the transmission corridor area in real time;

[0075] S52: Set appropriate alarm thresholds in the system. When the model identifies that the wildfire risk confidence level exceeds the threshold, the alarm mechanism is automatically triggered, and the time when the wildfire was identified and the location information of the transmission corridor area are sent to the transmission corridor security personnel via email and text messages, thereby ensuring the safety and stability of the power system in a timely and effective manner.

[0076] Compared with existing technologies:

[0077] The present invention extracts multi-level feature map information, so that the model can more accurately extract wildfire features, thereby effectively improving the accuracy of wildfire identification in complex transmission corridor environments. Through continuous learning and training, the accuracy and efficiency of wildfire identification can be further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a flow chart of the present invention;

[0079] Figure 2 A network diagram of a wildfire identification model for a power transmission corridor of the present invention;

[0080] Figure 3 This is the DFM structure diagram of the present invention;

[0081] Figure 4 This is the structural diagram of DG-SPPF of the present invention;

[0082] Figure 5 This is the structural diagram of AD-Head of the present invention. DETAILED DESCRIPTION

[0083] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0084] The present invention provides a method for identifying wildfires in power transmission corridors based on deep learning. Figure 1-Figure 5 , including the following specific steps:

[0085] S1: Construct a dataset of wildfire images in power transmission corridors;

[0086] Among them: The specific process of S1 is:

[0087] S11: Collect transmission corridor wildfire images or use the public transmission corridor wildfire image dataset to build a transmission corridor wildfire image dataset;

[0088] S12: manually label the collected transmission corridor wildfire image dataset, use the LabelImg image labeling tool to select the wildfire location in the transmission corridor wildfire image and label its category as wildfire;

[0089] In addition: labelImg is a tool for image data labeling. Its labeling method is to first select the location of the wildfire in the transmission corridor wildfire image, and then enter its category as wildfire;

[0090] S2: Feature enhancement of transmission corridor wildfire images in the transmission corridor wildfire image dataset;

[0091] Among them: The specific process of S2 is:

[0092] S21: Processing the transmission corridor wildfire image in the transmission corridor wildfire image dataset into an RGB image;

[0093] S22: Feature enhancement of wildfire features in the transmission corridor wildfire image dataset by increasing the R (red) element in the image and reducing the values ​​of the G (green) and B (blue) elements;

[0094] The processing calculation formula is as follows:

[0095] R_adj(x,y)=max(0,min(255,R_orig(x,y)×α))

[0096] G_adj(x,y)=max(0,min(255,G_orig(x,y)×β))

[0097] B_adj(x,y)=max(0,min(255,B_orig(x,y)×β))

[0098] Among them, R_orig(x,y), G_orig(x,y), and B_orig(x,y) represent the R, G, and B element values ​​of the original image at the position (x,y), ranging from 0 to 255. R_adj(x,y), G_adj(x,y), and B_adj(x,y) represent the R, G, and B element values ​​of the adjusted image at the position (x,y). α is used to increase the factor of the R element, which is a number greater than 1. β is used to reduce the factors of the G and B elements, which is a number between 0 and 1. Increasing the value of the R element can significantly enhance the red feature of the flame in the image, making it more prominent and easy to identify. At the same time, reducing the values ​​of the G and B elements helps to reduce the interference of the background and other non-flame areas, further highlight the flame area, and improve the contrast of the overall image. Then the image size is uniformly expanded or compressed to 512×512×3 to ensure the consistency of the size of the subsequent input images.

[0099] S3: A transmission corridor wildfire recognition model based on deep learning was built. The wildfire images of transmission corridors were input into the model, and the model automatically identified the wildfires in the transmission corridor images.

[0100] Among them: The specific process of S3 is:

[0101] S31: Design a DFM (Dynamic Perception Feature Extraction Module) dynamic perception feature extraction module, which expands the detail feature receptive field of the transmission corridor wildfire image through a dual-branch dynamic convolution operation, thereby better capturing the transmission corridor wildfire information in the image and enhancing the model's ability to identify transmission corridor wildfires, such as Figure 3 As shown;

[0102] Among them: The specific process of S31 is:

[0103] S311: Input the transmission corridor wildfire characteristic map T with size H×W×C 1 The convolution kernel number is C, the size is 3×3, and the padding value is 2. The transmission corridor wildfire feature map T with the size of H×W×C is obtained. 2 ;in:

[0104] The CBS module is a commonly used deep learning module, which consists of a Conv layer (convolution layer), a BN batch normalization layer, and a SiLu (activation function). It is mainly used for feature extraction and conversion. It extracts the features of the input data through convolution operations, uses the BN batch normalization layer for standardization, and provides nonlinear characteristics through the SiLu activation function.

[0105] An image of size H×W×C is an image with a length of H pixels, a width of W pixels, and C channels.

[0106] The number of channels of the feature map obtained after the image is calculated by the CBS module is the same as the number of convolution kernels in CBS;

[0107] S312: Transmission corridor wildfire characteristic map T with size H×W×C 2 , split along the channel to obtain the transmission corridor wildfire characteristic map T with a size of H×W×C / 2 3 and T 4 ;

[0108] S313: Transmission corridor wildfire characteristic map T 3 After the average pooling operation, the image is input into the dynamic convolution, followed by the BN batch normalization operation and the Swish activation function for nonlinear transformation. The result is then input into the GAM to obtain the transmission corridor wildfire feature map T with a size of H×W×C / 2. 5 ;

[0109] T 5 =GAM(Swish(BN(DConv(AVGPooling(T 3 )))))

[0110] AVGPooling stands for average pooling operation, DConv stands for dynamic convolution, BN stands for batch normalization, and GAM stands for global attention mechanism.

[0111] S314: Transmission corridor wildfire characteristic map T 4After the maximum pooling operation, it is input into the dynamic convolution, followed by the BN batch normalization operation and the Swish activation function for nonlinear transformation, and then input into the GAM to obtain the transmission corridor wildfire feature map T with a size of H×W×C / 2. 6 ;

[0112] T 6 =GAM(Swish(BN(DConv(MAXPooling(R 4 )))))

[0113] MAXPooling represents the maximum pooling operation, DConv represents dynamic convolution, BN represents batch normalization, and GAM represents the global attention mechanism. Average and maximum pooling operations, BN batch normalization, and Swish are all common operations in deep learning, while dynamic convolution and GAM global attention mechanism are existing technologies.

[0114] S315: Transmission corridor wildfire characteristic map T with size H×W×C / 2 5 and T 6 Superimpose along the channel to obtain the transmission corridor wildfire characteristic map T with a size of H×W×C 7 , T 7 Input into the CBS module with C convolution kernels, size 3×3, and padding value 2, and obtain and output the transmission corridor wildfire feature map T with size H×W×C. 8 ;

[0115] S32: Design a DF-SPPF (Dilated Convolution Fusion Spatial PyramidPooling Factorization) pyramid pooling structure so that the DF-SPPF pyramid pooling structure uses parallel dilated convolution branches to extract the features of the transmission corridor wildfire images, and fuses the features of the transmission corridor wildfire images of different scales through pooling operations, thereby improving the model's wildfire recognition ability for transmission corridor wildfire images of different scales, such as Figure 4 As shown;

[0116] Among them: The specific process of S32 is:

[0117] S321: Input transmission corridor wildfire feature map A with size H×W×C 1 In the first branch, A 1 After the calculation is performed with a dilated convolution with C convolution kernels, a size of 3×3, a padding value of 2, and an expansion rate of 2, the results are sequentially input into the BN batch normalization layer and the ReLu activation function to obtain a transmission corridor wildfire feature map A with a size of H×W×C. 2 , in the second branch A1 After the calculation of the dilated convolution with C convolution kernels, size 3×3, padding value 2, and expansion rate 4, the results are input into the BN batch normalization layer and ReLu activation function in turn to obtain the transmission corridor wildfire feature map A with size H×W×C. 3 , in the third branch A 1 After the calculation of the dilated convolution with C convolution kernels, size 3×3, padding value 2, and expansion rate 6, the results are input into the BN batch normalization layer and ReLu activation function in turn to obtain the transmission corridor wildfire feature map A with size H×W×C. 4 ;

[0118] A 2 =ReLU(BN(DLConv r＝2 (A 1 )))

[0119] A 3 =ReLU(BN(DLConv r＝4 (A 1 )))

[0120] A 4 =ReLU(BN(DLConv r＝6 (A 1 )))

[0121] Among them, DLConv r＝2 , DLConv r＝4 , DLConv r＝6 They represent the calculation of dilated convolution with expansion rates of 2, 4, and 6 respectively, BN represents batch normalization, and ReLU is the activation function;

[0122] S322: Transmission corridor wildfire feature map A with dimensions H×W×C 2 , A 3 , A 4 The transmission corridor wildfire characteristic map A with the size of H×W×3C is obtained by superimposing along the channel. 5 , then A 5 Input into the CBS module with C convolution kernels, size 3×3, and padding value 2, and obtain the transmission corridor wildfire feature map A with size H×W×C 6 ;

[0123] A 5 =Concat(A 2 , A 3 , A 4 )

[0124] Concat is a superposition operation along the channel;

[0125] S323: A 6 Input to three consecutive maximum pooling layers to obtain the transmission corridor wildfire feature map A with a size of H×W×C. 7 , A 8 , A 9 , the transmission corridor wildfire feature map A with the size of H×W×C 6 , A 7 , A 8 , A 9 The transmission corridor wildfire characteristic map A with the size of H×W×4C is obtained by superimposing along the channel. 10 , then A 10 Input into the CBS module with C convolution kernels, size 3×3, and padding value 2, and obtain and output the transmission corridor wildfire feature map A with size H×W×C 11 ;

[0126] A 7 =MAXPooling(A 6 )

[0127] A 8 =MAXPooling(A 7 )

[0128] A 9 =MAXPooling(A 8 )

[0129] A 10 =Concat(A 6 , A 7 , A 8、 A 9 )

[0130] S33: Design an AD-Head (Atrous Detection Head) hole detection head so that the AD-Head hole detection head can increase the receptive field of each convolution kernel while maintaining the size of the feature map, which helps to extract more contextual information from the input feature map and capture target features of different scales, so that the AD-Head hole detection head can extract rich feature information, thereby improving detection accuracy and efficiency, such as Figure 5 As shown;

[0131] Among them: The specific process of S33 is:

[0132] S331: Input the transmission corridor wildfire feature map L with size H×W×C 1 To the CBS module with C convolution kernels, size 3×3, and padding value 2, we get the transmission corridor wildfire feature map L with size H×W×C 2 ;

[0133] S332: L 2 Input into three branches. In the first branch, L 2 First, the transmission corridor wildfire feature map L with a size of H×W×C is obtained through a CBS module with a convolution kernel number of C, a size of 3×3, and a padding value of 2. 3 , then L 3 Input into a two-dimensional convolution layer with C convolution kernels and a size of 1×1 to obtain the H×W×C transmission corridor wildfire feature map L 4 , L 4 The target category of the characteristic image of wildfire in the power transmission corridor is obtained through calculation;

[0134] S333: In the second branch, L 2 The input is a dilated convolution with a kernel size of C, a size of 3×3, and an expansion rate of 2, and then BN batch normalization and ReLu activation function operations are performed to obtain a transmission corridor wildfire feature map L with a size of H×W×C. 5 , in the third branch, L 2 The input is a dilated convolution with a kernel size of C, a size of 3×3, and an expansion rate of 4, and then BN batch normalization and ReLu activation function operations are performed to obtain a transmission corridor wildfire feature map L with a size of H×W×C. 6 , L 5 and L 6 Add the features to get the transmission corridor wildfire feature map L with the size of H×W×C 7 , then L 7 Enter into two branches, in the upper branch L 7 The input is sent to a two-dimensional convolutional layer with 4 convolution kernels and a size of 1×1 to obtain the transmission corridor wildfire feature map L with a size of H×W×4. 8 , L 8 After calculating, the wildfire target regression frame of the wildfire feature image of the transmission corridor is obtained. In the lower branch, L 7 Input into a two-dimensional convolution layer with 1 convolution kernel and 1×1 size to obtain the transmission corridor wildfire feature map L with a size of H×W×1 9 , L 9 The wildfire target confidence of the wildfire feature image of the transmission corridor is obtained by calculation; wherein, the bounding box coordinate loss of the target of the present invention adopts SIoU loss, the confidence loss of the bounding box adopts cross entropy loss, and the classification loss of the wildfire adopts Focal Loss loss, and these three losses are all existing technologies;

[0135] S34: Using the DFM dynamic perception feature extraction module designed in S31, the DF-SPPF pyramid pooling structure designed in S32, and the AD-Head cavity detection head designed in S33, a transmission corridor wildfire recognition model is proposed. By inputting the transmission corridor wildfire image, the wildfire in the image is automatically recognized and located;

[0136] Among them: The specific process of S34 is:

[0137] S341: Input a transmission corridor wildfire image F with a size of 512×512×3 1 The convolution kernel number is 64, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 256×256×64 is obtained. 2 , and then F 2 Input into the CBS module with 128 convolution kernels, 3×3 size, 2 steps, and 1 padding value, and obtain the transmission corridor wildfire feature map F with a size of 128×128×128 3 ;

[0138] S342: F 3 The input convolution kernel number is 256, the size is 3×3, the step length is 2, and the padding value is 1. Then it is input into the DFM dynamic perception feature extraction module for feature extraction to obtain the transmission corridor wildfire feature map F with a size of 64×64×256. 4 , F 4 The input convolution kernel number is 512, the size is 3×3, the step size is 2, and the padding value is 1. Then it is input into the DFM dynamic perception feature extraction module for feature extraction to obtain the transmission corridor wildfire feature map F with a size of 32×32×512. 5 , F 5 Input the CBS module with 512 convolution kernels, 3×3 kernel size, 2 stride, and 1 padding value, and then input DFM to obtain the transmission corridor wildfire feature map F with a size of 16×16×512. 6 , F 6 Input into the DF-SPPF pyramid pooling structure to obtain the transmission corridor wildfire feature map F with a size of 16×16×512 7 ;

[0139] S343: F 7 The upsampling operation is performed to obtain the transmission corridor wildfire feature map F with a size of 32×32×512 8 , F 5 and F 8 After superimposing along the channel, the map is input into the DFM dynamic perception feature extraction module to obtain the transmission corridor wildfire feature map F with a size of 32×32×1024. 9 ;

[0140] S344: F 9 The upsampling operation is performed to obtain the transmission corridor wildfire feature map F with a size of 64×64×512 10 , F 4 and F 10 After superimposing along the channel, the map is input into the DFM dynamic perception feature extraction module to obtain the transmission corridor wildfire feature map F with a size of 64×64×768. 11 ;

[0141] S345: F 11 The number of input convolution kernels is 256, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 32×32×256 is obtained. 12 , F 9 and F 12 After superimposing along the channel, the map is input into the DFM dynamic perception feature extraction module to obtain the transmission corridor wildfire feature map F with a size of 64×64×1280. 13 ;

[0142] S346: Transmission Corridor Wildfire Characteristic Map F 13 The number of input convolution kernels is 512, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 32×32×512 is obtained. 14 , F 8 and F 14 After superimposing along the channel, the map is input into the DFM dynamic perception feature extraction module to obtain the transmission corridor wildfire feature map F with a size of 32×32×1024. 15 ;

[0143] S347: Transmission Corridor Wildfire Characteristic Map F 11 The number of input convolution kernels is 256, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 32×32×256 is obtained. 16 , the transmission corridor wildfire characteristic map F 13 The number of input convolution kernels is 512, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 16×16×512 is obtained. 17 , the transmission corridor wildfire characteristic map F 15 The number of input convolution kernels is 1024, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 8×8×1024 is obtained. 18 ;

[0144] S348: Transmission corridor wildfire feature map F16 、F 17 、F 18 The result is input into the AD-Head hole detection head to obtain and output the wildfire recognition result, which includes the target category, regression box and confidence of the wildfire in the image;

[0145] S4: Divide the transmission corridor wildfire image dataset, train and verify the transmission corridor wildfire recognition model;

[0146] Among them: The specific process of S4 is:

[0147] S41: Divide the transmission corridor wildfire image dataset constructed in S1 into a training set and a validation set at a ratio of 4:1, and train the transmission corridor wildfire recognition model constructed in S3;

[0148] S42: Initialize all neural network parameters and set model-related training parameters, divide the training set and validation set data into multiple batches, input a batch of training set data into the algorithm for training each time, and obtain the training loss value of the batch;

[0149] S43: Input the validation set into the model in batches to obtain the corresponding batch loss value, automatically learn and adjust the parameters according to the batch loss value each time, until the batch loss value tends to converge, and the transmission corridor wildfire recognition model training is completed;

[0150] S5: Apply the transmission corridor wildfire identification model;

[0151] Among them: The specific process of S5 is:

[0152] S51: After the training of the transmission corridor wildfire identification model is completed, the model is integrated into the existing transmission corridor monitoring system to ensure that the model can receive image or video data of the transmission corridor area in real time;

[0153] S52: Set appropriate alarm thresholds in the system. When the model identifies that the wildfire risk confidence level exceeds the threshold, the alarm mechanism is automatically triggered, and the time when the wildfire was identified and the location information of the transmission corridor area are sent to the transmission corridor security personnel via email and text messages, thereby ensuring the safety and stability of the power system in a timely and effective manner.

[0154] Example:

[0155] In this embodiment, 1565 transmission corridor wildfire images are collected for constructing a transmission corridor wildfire image dataset. The LabelImg image annotation tool is used to annotate the wildfire locations in the transmission corridor wildfire images with categories and coordinate frame information. Each image contains one or more wildfire targets, and the corresponding wildfire targets need to be selected with anchor frames multiple times. After all wildfire targets are labeled, 1565 annotation files in xml format can be obtained. The name of each annotation file is consistent with the name of the corresponding wildfire image file, and each annotation file can contain annotation anchor frame information and category information of multiple targets.

[0156] After the labeling is completed, the transmission corridor wildfire image in the transmission corridor wildfire image dataset is processed as an RGB image by increasing the R (red) element in the image and reducing the values ​​of the G (green) and B (blue) elements. Feature enhancement is performed on the wildfire features in the transmission corridor wildfire image dataset. The present invention uses python for processing. First, the transmission corridor wildfire image img is read in through the cv2.imread (img) of the opencv library, and then the b, g, r = cv2.split (img) function of the opencv library is used to separate the BGR channel of the transmission corridor wildfire image, where b, g, r represent the blue, green, and red channels of the transmission corridor wildfire image, respectively. Set the channel influence factor α = 1.3, β = 0.7, and use the numpy library r = np.clip (r × α, 0, 255). astype (np.uint8), g = np.clip (g × β, 0, 255). astype (np.uint8), b = np.clip (b × β, 0, 255). astype (np.uint8) to increase the R (red) element in the image and reduce the value of the G (green) and B (blue) elements. Use img_enhanced = cv2.merge ([b, g, r]) to merge the b, g, and r of the image, and then use img_reshape = img_enhanced.reshape (512, 512) to unify the size of the transmission corridor wildfire image to 512 × 512 × 3 (since the image is an RGB image, its number of channels is 3) and finally save the image cv2.imwrite (output_path, img_reshape). This operation is performed on each transmission corridor wildfire image in the transmission corridor wildfire image dataset until all processing is completed.

[0157] The constructed transmission corridor wildfire image data set is divided into a training set and a validation set at a ratio of 4:1, that is, 1252 images of the training set and 313 images of the transmission corridor wildfire in the test set. The transmission corridor wildfire recognition model designed in this scheme is used to train the feature-enhanced transmission corridor wildfire images to complete the construction of the transmission corridor wildfire recognition model. First, initialize the various parameters and hyperparameters of the algorithm training. This embodiment describes several important parameters in the algorithm training: the root initialized optimizer is the SGD optimizer, the initialization training epoch is 300, the batch_size is 32, and the initial learning rate is 0.001. These parameters need to be adaptively adjusted according to the results of multiple training of the algorithm until the algorithm modeling achieves the optimal effect.

[0158] After completing the initialization of the basic parameters of the algorithm training, the model training begins. The feature-enhanced power transmission corridor wildfire image is input into the power transmission corridor wildfire recognition model of the present invention. Assume that the input is a power transmission corridor wildfire image F with a size of 512×512×3 1 The convolution kernel number is 64, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 256×256×64 is obtained. 2 Then F 2 Input into the CBS module with 128 convolution kernels, 3×3 size, 2 steps, and 1 padding value, and obtain the transmission corridor wildfire feature map F with a size of 128×128×128 3 . 3 The input convolution kernel number is 256, the size is 3×3, the step size is 2, and the padding value is 1. Then it is input into the DFM for feature extraction to obtain the transmission corridor wildfire feature map F with a size of 64×64×256. 4 . 4 The input convolution kernel is 512, the size is 3×3, the step size is 2, and the padding value is 1. Then it is input into the DFM for feature extraction to obtain the transmission corridor wildfire feature map F with a size of 32×32×512. 5 . 5 Input the CBS module with 512 convolution kernels, 3×3 kernel size, 2 stride, and 1 padding value, and then input DFM to obtain the transmission corridor wildfire feature map F with a size of 16×16×512. 6 . 6 Input into DF-SPPF to obtain the transmission corridor wildfire feature map F with a size of 16×16×512 7 . 7 The upsampling operation is performed to obtain the transmission corridor wildfire feature map F with a size of 32×32×512 8 . 5 and F8 After superimposing along the channel, the input DFM is used to obtain the transmission corridor wildfire feature map F with a size of 32×32×1024. 9 . 9 The upsampling operation is performed to obtain the transmission corridor wildfire feature map F with a size of 64×64×512 10 . 4 and F 10 After superimposing along the channel, the input DFM is used to obtain the transmission corridor wildfire feature map F with a size of 64×64×768. 11 . 11 The number of input convolution kernels is 256, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 32×32×256 is obtained. 12 . 9 and F 12 After superimposing along the channel, the input DFM is used to obtain the transmission corridor wildfire feature map F with a size of 64×64×1280. 13 . Transmission corridor wildfire characteristic map F 13 The number of input convolution kernels is 512, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 32×32×512 is obtained. 14 . 8 and F 14 After superimposing along the channel, the input DFM is used to obtain the transmission corridor wildfire feature map F with a size of 32×32×1024. 15 . Transmission corridor wildfire characteristic map F 11 The number of input convolution kernels is 256, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 32×32×256 is obtained. 16 . Transmission corridor wildfire characteristic map F 13 The number of input convolution kernels is 512, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 16×16×512 is obtained. 17 . Transmission corridor wildfire characteristic map F 15 The number of input convolution kernels is 1024, the size is 3×3, the step size is 2, and the padding value is 1. The transmission corridor wildfire feature map F with a size of 8×8×1024 is obtained. 18 . Transmission corridor wildfire characteristic map F 16 、F 17 、F 18Input into AD-Head to obtain the target category, regression box, and confidence of the wildfire in the image. The training of the detection algorithm of the present invention performs parameter updates inside the algorithm through back propagation of the loss function. In the loss function described in the present invention, the bounding box coordinate loss of the target adopts SIoU loss, the confidence loss of the bounding box adopts binary cross entropy loss, and the classification loss for determining whether it is a wildfire adopts FocalLoss loss. The above loss function is updated round by round during the gradient back propagation process of algorithm training. The algorithm gradually converges in round by round training, and the detection accuracy gradually improves. After all rounds of training are completed, the algorithm will call the divided verification data to verify the algorithm's precision Box_Presice and recall Box_Recall, and mAP50 is used as the evaluation standard to test the effectiveness of algorithm training.

[0159] After the training of the transmission corridor wildfire identification model is completed, the model will be integrated into the existing transmission corridor monitoring system to ensure that the model can receive real-time image or video data of the transmission corridor area. The alarm threshold is set to 0.7 in the system. When the model identifies that the wildfire risk confidence exceeds the threshold, the alarm mechanism is automatically triggered, and the information such as the identified wildfire and the location of the transmission corridor area is sent to the transmission corridor security personnel via email and text messages, so as to ensure the safety and stability of the power system in a timely and effective manner.

[0160] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for identifying wildfires in power transmission corridors based on deep learning, characterized in that: The specific steps are as follows: S1: Construct a dataset of wildfire images in power transmission corridors; S2: Feature enhancement of transmission corridor wildfire images in the transmission corridor wildfire image dataset; S3: A transmission corridor wildfire recognition model based on deep learning was built. The wildfire images of transmission corridors were input into the model, and the model automatically identified the wildfires in the transmission corridor images. S4: Divide the transmission corridor wildfire image dataset, train and verify the transmission corridor wildfire recognition model; S5: Apply the transmission corridor wildfire identification model; The specific process of S3 is as follows: S31: Design a DFM dynamic perception feature extraction module; S32: Design a DF-SPPF pyramid pooling structure so that the DF-SPPF pyramid pooling structure can extract the features of the wildfire images in the power transmission corridor using parallel dilated convolution branches, and fuse the features of the wildfire images in the power transmission corridor at different scales through pooling operations; S33: Design an AD-Head void detection head; S34: Using the DFM dynamic perception feature extraction module designed in S31, the DF-SPPF pyramid pooling structure designed in S32, and the AD-Head cavity detection head designed in S33, a transmission corridor wildfire recognition model is proposed. By inputting the transmission corridor wildfire image, the wildfire in the image is automatically recognized and located; The specific process of S31 is as follows: S311: input the transmission corridor wildfire characteristic map T1 into the CBS module to obtain the characteristic map T2; S312: Split T2 along the channel to obtain feature maps T3 and T4; S313: Obtain feature map T5 by performing average pooling, dynamic convolution, BN batch normalization, Swish activation function and GAM on T3; S314: T4 is subjected to maximum pooling, dynamic convolution, BN batch normalization, Swish activation function and GAM to obtain feature map T6; S315: T5 and T6 are superimposed along the channel and input into the CBS module to obtain feature map T8; The specific process of S33 is as follows: S331: input the wildfire characteristic map L1 of the transmission corridor into the CBS module to obtain the characteristic map L2; S332: L2 is input into three branches. In the first branch, L2 is calculated through the CBS module and the two-dimensional convolution layer to obtain the target category. S333: In the second branch, L2 is input into the dilated convolution and then BN batch normalization and ReLu activation function are performed to obtain the feature map L5. In the third branch, L2 is input into the dilated convolution and then BN batch normalization and ReLu activation function are performed to obtain the feature map L6. The features of L5 and L6 are added to obtain the feature map L7. Then L7 is input into the two branches. In the upper branch, L7 is input into the two-dimensional convolution layer to calculate the regression box. In the lower branch, L7 is input into the two-dimensional convolution layer to calculate the target confidence.

2. According to a method for identifying forest fires in power transmission corridors based on deep learning according to claim 1, it is characterized in that: The specific process of S1 is as follows: S11: Collect transmission corridor wildfire images or use the public transmission corridor wildfire image dataset to build a transmission corridor wildfire image dataset; S12: The collected transmission corridor wildfire image dataset is manually labeled, and the LabelImg image labeling tool is used to select the wildfire location in the transmission corridor wildfire image and label its category as wildfire.

3. The method for identifying wildfires in power transmission corridors based on deep learning according to claim 1, characterized in that: The specific process of S2 is as follows: S21: Processing the transmission corridor wildfire image in the transmission corridor wildfire image dataset into an RGB image; S22: Enhance the wildfire features in the transmission corridor wildfire image dataset by increasing the R element in the image and reducing the values ​​of the G and B elements; The processing calculation formula is as follows: R_adj(x,y)=max(0,min(255,R_orig(x,y)×α)) G_adj(x,y)=max(0,min(255,G_orig(x,y)×β)) B_adj(x,y)=max(0,min(255,B_orig(x,y)×β)) Among them, R_orig(x,y), G_orig(x,y), and B_orig(x,y) represent the R, G, and B element values ​​of the original image at the (x,y) position, ranging from 0 to 255. R_adj(x,y), G_adj(x,y), and B_adj(x,y) represent the R, G, and B element values ​​of the adjusted image at the (x,y) position. α is used to increase the factor of the R element, which is a number greater than 1. β is used to reduce the factors of the G and B elements, which is a number between 0 and 1. Increasing the value of the R element can significantly enhance the red feature of the flame in the image, making it more prominent and easier to identify. At the same time, reducing the values ​​of the G and B elements can help reduce the interference of the background and other non-flame areas, further highlight the flame area, and improve the contrast of the overall image. Then the image size is uniformly expanded or compressed to 512×512×3 to ensure the consistency of the size of the subsequent input images.

4. The method for identifying wildfires in power transmission corridors based on deep learning according to claim 1, characterized in that: The specific process of S4 is as follows: S41: Divide the transmission corridor wildfire image dataset constructed in S1 into a training set and a validation set at a ratio of 4:1, and train the transmission corridor wildfire recognition model constructed in S3; S42: Initialize all neural network parameters and set model-related training parameters, divide the training set and validation set data into multiple batches, input a batch of training set data into the algorithm for training each time, and obtain the training loss value of the batch; S43: Input the validation set into the model in batches to obtain the corresponding batch loss value, automatically learn and adjust the parameters according to the batch loss value each time, until the batch loss value tends to converge, and the training of the transmission corridor wildfire recognition model is completed.

5. The method for identifying wildfires in power transmission corridors based on deep learning according to claim 1, characterized in that: The specific process of S5 is as follows: S51: After the training of the transmission corridor wildfire identification model is completed, the model is integrated into the existing transmission corridor monitoring system to ensure that the model can receive image or video data of the transmission corridor area in real time; S52: Set appropriate alarm thresholds in the system. When the model identifies that the wildfire risk confidence level exceeds the threshold, the alarm mechanism is automatically triggered, and the time when the wildfire was identified and the location information of the transmission corridor area are sent to the transmission corridor security personnel via email and text messages, thereby ensuring the safety and stability of the power system in a timely and effective manner.

Images