Power transmission line forest fire detection method based on KAN-YOLO network

Abstract

The invention discloses a power transmission line forest fire detection method based on a KAN-YOLO network, and relates to the technical field of intelligent power grid operation and maintenance. The detection method comprises the following steps: constructing a power transmission line forest fire detection model based on a KAN-YOLO network, and inputting a power transmission line image into the model to obtain a forest fire event reasoning result; the model construction method comprises the following steps: acquiring forest fire accident video data, segmenting and classifying the data to obtain an image data set for network training, and dividing the image data set into a training set, a verification set and a test set; preprocessing the data set; constructing a power transmission line forest fire detection model based on a KAN-YOLO network; training the model by using the training set, and checking a training result by using the verification set to obtain an intermediate training model; and performing performance evaluation on the intermediate training model by using the test set, and selecting the model with the optimal performance. Compared with a traditional YOLO model, the model provided by the invention greatly improves the recognition precision of forest fire detection, reduces the false detection rate, improves the efficiency of power transmission line maintenance, operation and maintenance, and is high in practicability.

Description

Technical Field

[0001] The present invention relates to the field of smart grid operation and maintenance technology, and in particular to a transmission line wildfire detection method based on a KAN-YOLO network. Background Art

[0002] Frequent wildfires have become an important factor affecting the safety and stability of power transmission lines. Wildfires may not only cause failures, short circuits, or even fires in transmission lines, but may also cause serious losses to the power system, affecting the reliability and safety of power supply. Traditional wildfire monitoring methods mainly rely on manual inspections and meteorological early warning systems, but these methods have obvious limitations. Manual inspections require a lot of manpower and material resources, are inefficient, and pose safety hazards; and although the meteorological early warning system can provide certain early warning information, it is often difficult to accurately predict the location and time of wildfires, and there are problems such as narrow monitoring range and poor real-time performance. With the continuous development and application of sensor technology, drone technology, and artificial intelligence technology, new possibilities have been provided for wildfire detection in transmission lines.

[0003] Compared with traditional wildfire detection methods, convolutional neural networks (CNN) in the field of deep learning have performed well in image recognition, classification, and detection. The YOLO series of algorithms are built based on CNN, which can automatically learn and extract effective features from large amounts of image data. The YOLO algorithm is a real-time target detection algorithm that converts the target detection task into a single forward pass regression problem, significantly improving the detection speed and accuracy. However, directly applying YOLO to transmission line wildfire detection also faces some challenges: such as high model complexity, high reliance on manual experience, and poor knowledge transferability.

[0004] To solve the above problems, the applicant introduced KAN to improve the YOLOv5 algorithm. The KAN network based on the Kolmogorov-Arnold theorem applies a learnable activation function to the weights instead of using a fixed activation function on the nodes (neurons). Therefore, the KAN network can process and learn the complex relationships of the input data more flexibly and accurately. This makes the KAN network have higher flexibility and expressiveness in function approximation and complex data modeling. The KAN-YOLO network has significant effects and advantages in improving the flexibility and expressiveness of the transmission line wildfire detection model, maintaining the interpretability of the model, reducing the number of parameters and improving performance, and enhancing scalability. Summary of the invention

[0005] In view of the above shortcomings, the present invention provides a power transmission line wildfire detection method based on the KAN-YOLO network, which can improve the flexibility and expression ability of the power transmission line wildfire detection model, reduce the number of parameters and improve the detection performance, and enhance the scalability of the method. The specific technical solution is as follows:

[0006] A method for detecting wildfires in power transmission lines based on a KAN-YOLO network, the method comprising: constructing a wildfire detection model for power transmission lines based on a KAN-YOLO network, inputting a real-time collected power transmission line image into the wildfire detection model for power transmission lines based on the KAN-YOLO network for wildfire detection, and obtaining a position of the wildfire accident relative to the input image and the level of the accident; the method for constructing the wildfire detection model comprises the following steps:

[0007] S101: Obtain video data of wildfire accidents on transmission lines, and construct a data set of wildfire transmission lines on transmission lines; extract image frames of wildfire accidents from the power grid transmission line environment video collected by the inventor using equipment such as drones. Classify images according to different environments, different accident levels, different image frames and other conditions. Use image annotation tools to annotate the location of the wildfire accident and the level of the wildfire accident. Obtain an image data set that can be used for network training. After constructing the data set, use geometric data augmentation, add noise and other methods to further expand the data set, and then divide the expanded data set into training set, validation set and test set in a ratio of 7:2:1;

[0008] S102: Before the data set constructed in S101 is fed to the provided network method for training, it needs to be preprocessed;

[0009] S103: Construct a transmission line wildfire detection model based on KAN-YOLO network;

[0010] S104: using the training set preprocessed by S102 to train the power transmission line wildfire detection model constructed by S103, and using the validation set preprocessed by S102 to verify the training result, to obtain an intermediate training model;

[0011] S105: Use the test set processed by S102 to perform performance evaluation on the intermediate training model obtained in S104, and select the power transmission line wildfire detection model with the best performance, which is the power transmission line wildfire detection model based on the KAN-YOLO network.

[0012] Furthermore, in step S102, the preprocessing includes: Mosaic data enhancement, adaptive anchor frame calculation and adaptive image scaling.

[0013] Furthermore, in step S103, the power line wildfire detection model based on the KAN-YOLO network is based on an improved YOLOv5 variant, including: an input end, a backbone network, a Neck network and an output end.

[0014] Furthermore, the backbone network includes: a Focus module, a C3 module of the backbone network and an SPPF module, which are responsible for extracting the image features fed from the input end;

[0015] The Focus module first slices the input image (length × width × number of channels) in the format of 640 × 640 × 3 to obtain four sub-images of 320 × 320 × 3, and then splices the four sub-images in the channel dimension to obtain a feature map of 320 × 320 × 64; finally, a convolution operation of size 3 × 3 and stride 1 is selected to achieve downsampling of the feature map and adjustment of the number of channels;

[0016] The CSP module and the KAN-BottleNeck module are nested in the C3 module of the backbone network; the CSP module adopts the CSP-Darknet53 structure, specifically, the CSP module divides the input feature map into two branches for parallel processing, one branch performs a conventional convolution operation, and the other branch directly performs a splicing operation, and finally, the results of the two branches are merged, the CSP module not only retains the diversity of image feature information, but also reduces the amount of calculation, making it more convenient to deploy on low-computing power devices; based on the combination of CSP modules, a C3 structure is constructed, the CSP module and the BottleNeck module are nested in the C3 structure, the correction unit used in the C3 module is SiLU, the present invention improves the BottleNeck in the backbone network, specifically, the BottleNeck block is replaced by the KAN-BottleNeck module, the KAN-BottleNeck module applies a learnable activation function to the weight based on the KAN network of the Kolmogorov-Arnold theorem, instead of using a fixed activation function on the node (neuron);

[0017] The SPPF module: The fast spatial pyramid pooling module (Spatial Pyramid Pooling–Fast, SPPF) is a middleware connecting the backbone network and the Neck network. Specifically, after the data from the backbone network undergoes a convolution operation on one side, it is branched based on the CSP module, one of which uses three 5×5 maximum pooling layers to operate the feature map, and finally the results of the three pooling layers are spliced ​​with the other branch, and then a layer of convolution operation is performed on the input feature map for multi-scale fusion to enhance the expressiveness of the feature map.

[0018] Furthermore, the Neck network is mainly composed of multiple convolution operations, upsampling operations, feature connections and the C3 module of the Neck network; it adopts a feature pyramid network (FPN) and a path aggregation network structure (PAN) to achieve multi-scale feature fusion, wherein the feature pyramid network FPN splices shallow image features with deep image semantic features through upsampling operations and lateral connections of feature maps to improve the representation ability of feature maps; the path aggregation network PAN adds bottom-up path enhancement on the basis of the feature pyramid network FPN, further improving the effect of feature fusion.

[0019] Furthermore, the C3 module in the Neck network is different from the C3 module in the backbone network in that the C3 module in the Neck network does not use the KAN-Conv operation in the BottleNeck structure and does not directly concatenate the intermediate features.

[0020] Furthermore, the output end is the prediction part of KAN-YOLO, which is used to generate the results of power line wildfire detection; the output end uses the CIoU loss function to calculate the difference between the predicted box and the real box. When generating the final detection result, KAN-YOLO uses non-maximum suppression (NMS) processing to eliminate redundant detection results.

[0021] Furthermore, in step S104, the training is to read the prepared training data set in the experimental environment, and use the corresponding deep learning development platform to perform multiple batches and multiple rounds of training based on minimizing the following losses, and cross-use the validation set to evaluate the model performance during the training process, and take the model with the best performance in the training process as the result model.

[0022] Furthermore, the power transmission line wildfire detection model based on the KAN-YOLO network can be deployed in low-computing power devices to perform real-time reasoning on the power transmission line images collected during the inspection process to generate wildfire event reasoning results.

[0023] Furthermore, the low computing power device is a drone.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] Compared with the YOLOv5 model before improvement, the KAN-YOLO network of the present invention has significant effects and advantages in improving the flexibility and expression ability of the power transmission line wildfire detection model, maintaining the interpretability of the model, reducing the number of parameters and improving performance, and enhancing scalability, reducing the false detection rate, improving the efficiency of power transmission line inspection and maintenance, and facilitating deployment on low-computing power equipment. It has strong practicality and great application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for describing the embodiments are briefly introduced below.

[0027] Figure 1 is a basic flow chart of the example method of the present invention;

[0028] Figure 2 It is an architecture diagram of a power transmission line wildfire detection method based on a KAN-YOLO network proposed in the present invention;

[0029] Figure 3 It is a structural diagram of the KAN-BottleNeck module proposed in the present invention;

[0030] Figure 4 It is the BottleNeck module structure diagram of the present invention and the C3 module structure diagram in the Neck network;

[0031] Figure 5 It is the structural diagram of the SPPF module in the present invention. DETAILED DESCRIPTION

[0032] The specific embodiments of the present invention are described in detail below, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

[0033] Example 1

[0034] As attached Figure 1 As shown, this embodiment provides a method for detecting forest fires on power transmission lines based on a KAN-YOLO network, comprising the following steps:

[0035] S101, construct a transmission line wildfire detection image dataset.

[0036] The data set in this example is from videos of several wildfire incidents in mountainous areas under the power transmission lines in a certain network area collected by the inventor using drones and other equipment (preliminary blurring has been performed), and image frames of wildfire accidents are extracted from these videos. Images are classified according to different environments, different accident levels, different image frames and other conditions. The image annotation tool used is LabelMe, which is used to annotate the location and level of wildfire accidents. An image data set that can be used for network training is obtained. After the data set is constructed, the data set is further expanded using methods such as geometric data augmentation and adding noise, and then the expanded data set is divided into training set, validation set, and test set in a ratio of 7:2:1.

[0037] S102, establishing a power transmission line wildfire detection model based on the KAN-YOLO network.

[0038] The details of the model flow established in S102 for the transmission line wildfire image data are attached. Figure 2 .

[0039] The established model includes: feature extraction backbone network, Neck network, and output end / prediction head network.

[0040] The backbone network in this example includes: Focus module, CSP module, SPPF module; the CSP module is constructed using the CSP-DarkNet53 network structure. The SPPF module is improved based on the SPP module and has higher efficiency.

[0041] Based on the combination of CSP modules, a C3 structure is constructed. CSP modules and BottleNeck modules are nested in the C3 structure. The correction unit used in the C3 module is SiLU. The present invention improves BottleNeck in the backbone network. Specifically, the BottleNeck1 block is replaced by the KAN-BottleNeck module; the KAN-BottleNeck module is based on the Kolmogorov-Arnold theorem KAN network, which applies a learnable activation function to the weights instead of using a fixed activation function on the nodes (neurons); the basic principles of the KAN network are as follows:

[0042] The Kolmogorov-Arnold representation theorem assumes that a multivariate function f can be represented as:

[0043]

[0044] Among them, φ q,p is the mapping input variable x p A single variable function, φ q,p The subscripts of are the sub-terms of the summation term. KAN converts each layer

[0045] The structure is represented as a set of learnable one-dimensional function matrices:

[0046] Φ={φ q,p},p=1,2,K n in ,q=1,2,K n out

[0047] In particular, each φ q,p Functions can be defined as B-curves, which are spline functions defined by linear combinations of basis curves, enhancing the network's ability to learn complex data representations; in represents the number of input features of a particular layer, and n out Indicates the number of output features generated by this layer, reflecting the dimensionality conversion between network layers. The activation function φ in the matrix l,j,i That is, a learnable zone equation, expressed as follows:

[0048]

[0049] Among them, c i is a trainable coefficient. The above formula allows each φ l,j,i Adapting its shape based on the data provides unprecedented flexibility in how the network models interactions between inputs.

[0050] The overall structure of KAN is similar to the stacked layers in a multi-layer perceptron, but it is enhanced by using complex function mapping instead of simple linear transformation and nonlinear activation. See the following formula:

[0051] KAN(x)=(Φ L-1 oΦ L-2 oLΦ0)(x)

[0052] For each layer of transformation: Φ l The input x of the lth layer l Mapped to the input x of the next layer l+1 The specific process is as follows:

[0053]

[0054] Where each activation function φ l,j,i is a spline that provides a rich and adaptable response to the model input.

[0055] Based on KAN, this embodiment proposes a KAN-Conv operation, which is similar to the traditional CNN operation, but the difference is that the convolution layer Conv is replaced by KAN-Conv. Conv operation is generally connected to a layer of multi-layer perceptron after the feature map is flattened to learn it. KAN-Conv is connected to the KAN layer after the feature map is flattened. The advantage of KAN-Conv is that it requires much fewer parameters. This is determined by the structure of the network, because the B-curve can smoothly represent any activation function, that is, it does not need to perform a linear correction operation.

[0056] The main difference between the convolution used in KAN-Conv and Conv is the convolution kernel. In Conv, it is composed of weights, while in KAN-Conv, each element in the convolution kernel, φ, is a learnable nonlinear function that is learned using the B-curve. See the following formula:

[0057] φ=ω1·spline(x)+ω2·SiLU(x)

[0058] Among them, ω1 and ω2 are learnable parameters in the network, representing the network weights; Spline represents the spline curve about x, and SiLU is an improved activation function of Sigmoid and ReLU.

[0059] In KAN-Conv, the convolution kernel slides over the image and converts the corresponding activation function φ i,j Applied to the corresponding pixel a k,l , and the output pixel is calculated as φ i,j (a k,l ). Let K be the sum of N×M The KAN-Conv kernel on the input image is expressed as the following matrix:

[0060]

[0061] Furthermore, KAN-Conv can be defined as follows:

[0062]

[0063] Image represents image, a represents pixel, and the i and j subscripts are the x and y coordinates of the pixel points (assuming it is two-dimensional).

[0064] Supplementary Example 1: A KAN-Conv kernel of size 2×2 is as follows:

[0065]

[0066] The calculation process of formula 1 is as follows:

[0067]

[0068] Based on the KAN network, this embodiment proposes a KAN-BotttleNeck module, namely Figure 2 The C3 module marked in the backbone network has a workflow shown in Figure 3 Specifically, the image features are first subjected to a conventional convolution operation, and then the convolution result is subjected to a feature map slicing operation, which is divided into four branches, three of which enter the proposed KAN-Conv convolution operation. These three branches are cascaded in sequence, and the intermediate image features are subjected to KAN-Conv convolution to obtain three intermediate feature map results. Then, these three intermediate features are spliced ​​with the last branch of the feature map slice, and finally, the above feature map splicing results are subjected to a conventional convolution operation again. The role of this module is to reduce the number of module parameters while learning richer image semantic features, so that image features can be processed more efficiently and accurately;

[0069] The SPPF module is an improved variant based on the SPP module. Figure 5 , where the ConvBNSiLU module is a combination of convolution operation, batch normalization layer, and linear correction unit SiLU. Its working principle is to perform ConvBNSiLU operation on the feature map, and then perform branch operation on the feature map based on the residual idea. One branch performs three maximum pooling operations on the feature map in sequence, and the results of these three maximum pooling operations are spliced ​​with another feature branch, and finally perform another ConvBNSiLU operation. The purpose is to obtain image features at different levels and enhance the representation effect of image features.

[0070] In summary, the flow of the transmission line wildfire image in the embodiment of the backbone network is as follows:

[0071] 1. Input an input image with a length × width × number of channels of 640 × 640 × 3 into the Focus module to obtain a feature map of 160 × 160 × 128, which is then sent to the C3 module for learning. It is then connected to the ConvBNSiLU module to obtain a feature map of 80 × 80 × 256.

[0072] 2. The feature map obtained in 1. is then connected to the C3 module and the ConvBNSiLU module to learn features, and a feature map of size 40×40×512 is obtained. Among them, the C3 module generates two branches, one branch is passed to the next layer of the backbone network, and the other branch is passed to the Neck network for feature connection.

[0073] 3. The feature map obtained in 2. is connected to the C3 module and SPPF module for the final step of learning, and a 20×20×1024 feature map result is obtained. Among them, the C3 module generates two branches, one branch is passed to the lower layer KAN-SPPF module, and the other branch is passed to the Neck network for feature connection.

[0074] The Neck network in this embodiment includes: a ConvBNSiLU layer, an upsampling layer, a feature connection layer, and a C3 module in the Neck network.

[0075] The C3 module in the Neck network is different from the C3 module in the backbone network. The difference is that the KAN-Conv operation is not used in the BottleNeck structure, and the intermediate features are not directly spliced. The C3 module structure in the Neck network and the BottleNeck structure are shown in the attached Figure 4 .

[0076] The Neck network is built based on the FPN and PAN network architecture ideas. The process is as follows:

[0077] 1) Receive the features from the last SPPF layer in the backbone network, perform ConvBNSiLU operations on them, and obtain a 20×20×512 feature map. The feature map is divided into two branches for transmission, one branch is passed to the second-stage feature concatenation branch, and the other branch is passed to the lower network for upsampling. Finally, it is concatenated with the 40×40×512 features passed from the backbone network process 3), and the concatenation result is passed to the C3 module.

[0078] 2.) Perform a ConvBNSiLU operation on the feature map generated in 1) to obtain a 40×40×256 feature map. Based on the PAN network concept, the feature map is divided into two branches for transmission, one branch is passed to the lower network, and the other branch is passed to the second-stage feature concatenation through a step connection.

[0079] 3) Perform an upsampling operation on the feature map generated in 2) to obtain a feature map of 80×80×256, and then concatenate the feature map with the feature map from the backbone network process 2) and send it to the C3 module. The resulting feature result of 80×80×256 is divided into two branches for transmission. One branch is directly sent to the two-dimensional convolution layer as the prediction head input 1 at the output end. The other branch performs the ConvBNSiLU operation to obtain a 40×40×256 feature map result. The result is then concatenated with the feature map of the two-stage feature concatenation branch in 2), and finally the feature map is sent to the C3 module. The C3 module is divided into two branches for transmission, one branch is passed to the lower network, and the other branch is passed to the two-dimensional convolution layer as the prediction head input 2 at the output end;

[0080] 4) After performing the ConvBNSiLU operation on the result obtained in 3), a 20×20×512 feature map is obtained, and the feature map is connected with the feature map of the two-stage feature concatenation branch from 1), and the connected feature map is sent to the C3 module. Finally, the generated result is sent to the two-dimensional convolution layer as the prediction head input 3.

[0081] The output / prediction head in this example contains three inputs, which is the prediction part of KAN-YOLO, used to generate the results of power line wildfire detection. It uses the CIoU loss function to calculate the difference between the predicted box and the true box. In addition, when generating the final detection result, KAN-YOLO uses NMS processing to eliminate redundant detection results;

[0082] S104, read the prepared training data set in the experimental environment, use the corresponding deep learning development platform to perform multiple batches and multiple rounds of training based on minimizing the following losses, and cross-use the validation set to evaluate the model performance during the training process, and take the model with the best performance during the training process as the result model.

[0083]

[0084] CIoU (Complete Intersection over Union) is a metric used to evaluate the similarity between the predicted bounding box and the true bounding box in object detection tasks.

[0085] This embodiment relies on the following conditions, see Table 1 for training experiments.

[0086] Table 1

[0087]

[0088] In order to accelerate network convergence, the network optimizer used in this embodiment is Adam, and the initial learning rate is set to 1×10 3 .

[0089] S105, evaluate the intermediate model trained in S104, and select the model with the best evaluation performance as the final model. In order to accelerate network convergence, this example uses the cross-validation method during training to evaluate the intermediate model. Specifically, after every 200 training processes, the overall model is frozen, and 50 training processes are verified on the basis of freezing, and then the model is unfrozen, and this is repeated. Until 5000 training processes are trained. Since the image size in the data set is fixed, a large amount of graphics card memory is consumed. In this example, the batch size of each training process is set to 8.

[0090] S106, applying the final model selected in S105 to complex scenarios in practice. Specifically, the final model is regarded as a method for detecting wildfires in power transmission lines based on a KAN-YOLO network as proposed in the present invention. The model is deployed in low-computing-power devices such as drones, and real-time reasoning is performed on the power transmission line images collected during the inspection process to generate wildfire event reasoning results. Assist technical experts to efficiently complete inspection tasks.

[0091] The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.

Claims

1. A method for detecting forest fires on power transmission lines based on a KAN-YOLO network, characterized in that: The detection method comprises: constructing a power transmission line wildfire detection model based on a KAN-YOLO network, inputting the power transmission line image collected in real time into the power transmission line wildfire detection model based on the KAN-YOLO network, and obtaining the location and level of the wildfire accident; The method for constructing the mountain fire detection model comprises the following steps: S101: Obtain transmission line wildfire accident video data with known wildfire locations and accident levels, segment the video data into image frames, and classify the images according to different environments, different accident levels, and different image frames to obtain an image data set that can be used for network training. After the data set is constructed, further expand the data set and divide the expanded data set into a training set, a validation set, and a test set in a ratio of 6-8:1-3:1; S102: preprocessing the training set, validation set, and test set constructed in S101; S103: Construct a transmission line wildfire detection model based on KAN-YOLO network; S104: using the training set preprocessed by S102 to train the power transmission line wildfire detection model constructed by S103, and using the validation set preprocessed by S102 to verify the training result, to obtain an intermediate training model; S105: Use the test set processed by S102 to perform performance evaluation on the intermediate training model obtained in S104, and select the power transmission line wildfire detection model with the best performance, which is the power transmission line wildfire detection model based on the KAN-YOLO network.

2. According to claim 1, a method for detecting forest fires on power transmission lines based on a KAN-YOLO network is characterized in that: In step S102, the preprocessing includes: Mosaic data enhancement, adaptive anchor frame calculation and adaptive image scaling.

3. According to claim 1, a method for detecting forest fires on power transmission lines based on a KAN-YOLO network is characterized in that: In step S103, the power transmission line wildfire detection model based on the KAN-YOLO network includes: an input end, a backbone network, a Neck network and an output end.

4. According to claim 3, a method for detecting forest fires on power transmission lines based on a KAN-YOLO network is characterized in that: The backbone network includes: a Focus module, a C3 module of the backbone network and an SPPF module, which are responsible for extracting the image features fed from the input end; The Focus module first slices the input image to obtain four sub-images, then splices the sub-images in the channel dimension to obtain a feature map, and finally selects a convolution operation of a certain size and step length to achieve downsampling of the feature map and adjustment of the number of channels; The CSP module and the KAN-BottleNeck module are nested in the C3 module of the backbone network; the CSP module adopts the CSP-Darknet53 structure. Specifically, the CSP module divides the input feature map into two branches for parallel processing, one branch performs a conventional convolution operation, and the other branch directly performs a splicing operation. Finally, the results of the two branches are merged. The KAN-BottleNeck module applies a learnable activation function to the weights based on the KAN network of the Kolmogorov-Arnold theorem, instead of using a fixed activation function on the nodes (neurons); The SPPF module is a connector between the backbone network and the Neck network. After a layer of convolution operation, the data from the backbone network is branched based on the CSP module. One branch uses three maximum pooling layers to operate the feature map. Finally, the results of the three pooling layers are spliced ​​with another branch, and then a layer of convolution operation is performed to perform multi-scale fusion on the input feature map.

5. The method for detecting forest fires on power transmission lines based on a KAN-YOLO network according to claim 3, characterized in that: The Neck network is mainly composed of multiple convolution operations, upsampling operations, feature connections and the C3 module of the Neck network; it adopts a feature pyramid network and a path aggregation network structure, wherein the feature pyramid network splices shallow image features with deep image semantic features through upsampling operations and lateral connections of feature maps; the path aggregation network adds bottom-up path enhancement on the basis of the feature pyramid network.

6. A method for detecting forest fires on power transmission lines based on a KAN-YOLO network according to claim 5, characterized in that: The C3 module in the Neck network is different from the C3 module in the backbone network. The difference is that the C3 module in the Neck network does not use the KAN-Conv operation in the BottleNeck structure and does not directly concatenate the intermediate features.

7. A method for detecting forest fires on power transmission lines based on a KAN-YOLO network according to claim 3, characterized in that: The output end is the prediction part of KAN-YOLO, which is used to generate the results of power line wildfire detection; the output end uses the CIoU loss function to calculate the difference between the predicted box and the real box. When generating the final detection result, KAN-YOLO uses non-maximum suppression processing to eliminate redundant detection results.

8. The method for detecting forest fires on power transmission lines based on a KAN-YOLO network according to claim 1, characterized in that: In step S104, the training is to read the prepared training data set in the experimental environment, and use the corresponding deep learning development platform to perform multiple batches and multiple rounds of training based on minimizing the following losses, and cross-use the validation set to evaluate the model performance during the training process, and take the model with the best performance in the training process as the result model.

9. The method for detecting forest fires on power transmission lines based on a KAN-YOLO network according to claim 1, characterized in that: The power transmission line wildfire detection model based on the KAN-YOLO network can be deployed in low-computing-power devices to perform real-time reasoning on the power transmission line images collected during the inspection process to generate wildfire event reasoning results.

10. A method for detecting forest fires on power transmission lines based on a KAN-YOLO network according to claim 9, characterized in that: The low computing power device is a drone.

Images