Treeline fault identification and early warning method based on DCN-YOLO model

Abstract

The present invention discloses a tree-line fault identification and early warning method based on the DCN-YOLO model, comprising: simulating real tree-line discharges through a tree-line discharge experimental platform, shooting videos and exporting pictures frame by frame, and labeling the pictures to obtain an initial data set; obtaining flame, cable, and branch data sets to expand the initial data set to form an enhanced data set, and dividing the data set into a training set, a validation set, and a test set; performing data enhancement, improving the channel attention mechanism, and improving the sample weight balance of the loss function based on the DCN-YOLO model; inputting the enhanced data set into the improved model for training and verification, obtaining the highest evaluation indexes mAP50 and mAP50-90 values, and obtaining a tree-line fault monitoring and early warning model; connecting the tree-line fault monitoring and early warning model to a transmission tower monitoring system, and triggering an alarm system when branches, flames, or either of the two are detected on the transmission line. The present invention realizes accurate early warning of tree-line fault fires in the early stages.

Description

Technical Field

[0001] The present invention relates to the technical field of power disaster prevention, and in particular to a tree line fault identification and early warning method based on a DCN-YOLO model. Background Art

[0002] Model detection for treeline faults is based on image processing and deep learning technologies. Transmission lines are crucial energy transmission facilities in power systems. Treeline contact or discharge faults in transmission lines passing through forests can cause forest fires or power line tripping. Rapidly identifying and warning treeline faults and taking effective measures to prevent them can effectively improve power system stability and reduce the probability of forest fires. Currently, there is a lack of target detection for treeline faults. Deep learning-based target detection algorithms offer greater robustness and accuracy than traditional feature-based target detection algorithms. The YOLO series, with its fast computational speed and ease of training, can better meet the real-time requirements of fault identification and warning.

[0003] Existing research on transmission lines primarily focuses on identifying images of wildfires beneath the lines. This technology fails to accurately provide early warnings of fires caused by tree-line faults on the lines, thereby mitigating the likelihood of forest fires caused by these faults. Currently, when a tree-line fault occurs due to three-phase voltage imbalance in the power grid, it is difficult to accurately identify the faulty line. This results in an inability to effectively and timely shut down the faulty line, impacting the safety and stability of the power system. Summary of the Invention

[0004] In order to address the shortcomings of existing image recognition technology, the present invention provides a tree-line fault identification and early warning method based on the DCN-YOLO model. Aiming at the problem that the current transmission line image recognition technology is lacking in tree-line fault identification technology, an improved DCN-YOLO model is proposed in the existing target detection technology. By enhancing the data set and adjusting the improved attention mechanism and loss function parameters, a tree-line fault monitoring and early warning model is obtained and connected to the early warning system, realizing real-time online identification, monitoring and early warning of transmission line tree-line faults.

[0005] To achieve the above objectives, the present invention adopts a technical solution: a tree line fault identification and early warning method based on the DCN-YOLO model, comprising the following steps:

[0006] Step 1: Use the treeline discharge experimental platform to simulate real treeline discharges, shoot videos, export images frame by frame, and label the images to obtain the initial dataset;

[0007] Step 2: Obtain the flame, cable, and tree branch datasets to expand the initial dataset to form an enhanced dataset, and divide it into a training set, a validation set, and a test set;

[0008] Step 3: Based on the DCN-YOLO model, perform data enhancement, improve the channel attention mechanism, and improve the loss function sample weight balance;

[0009] Step 4: Input the enhanced dataset into the improved model for training and verification, obtain the highest evaluation index mAP50 and mAP50-90 values, and obtain the tree line fault monitoring and early warning model;

[0010] Step 5: Connect the tree-line fault monitoring and early warning model to the transmission tower monitoring system. When branches, flames, or either of the two are detected on the transmission line, the alarm system will be triggered.

[0011] As a further improvement of the present invention, in step 1, image data when a tree line fault occurs is obtained by using image acquisition equipment installed around the transmission line, and the image acquisition equipment includes a movable camera and a visual monitoring device set at a fixed node of the tower; the initial data set refers to the data set of flame, cable, and branch labels after labeling.

[0012] As a further improvement of the present invention, the step 3 specifically includes the following steps:

[0013] Step 3.1. Build the DCN-YOLO model, which includes the backbone network for extracting features from the input image, the head network for making decisions and generating the final detection results, and the neck network for feature fusion and enhancement. The deformable convolutional layer DCNv3 is introduced to improve the performance of handling complex deformable targets.

[0014] Step 3.2: Randomly adjust the color attributes of the dataset to enhance the DCN-YOLO model's ability to recognize targets under different lighting and color conditions. Also, randomly rotate the image by up to 15 degrees to simulate different possible orientations of the target in the image. Randomly crop a region from the original dataset image and resize it to 224x224 for the DCN-YOLO model to learn targets of different sizes and proportions. Apply Gaussian blur to the dataset image to simulate images at different focal lengths and enhance the DCN-YOLO model's invariance to details.

[0015] Step 3.3: Integrate a channel attention mechanism into the DCN-YOLO model. Use global average pooling and global maximum pooling to obtain global information of the feature map. Then, use two 1×1 convolutional fully connected layers and ReLU activation functions to learn the weight relationship between channels. Finally, use the Sgmoid function to compress the output to the [0, 1] interval as the channel attention weight.

[0016] Step 3.4: Improve the DCN-YOLO model through the focal loss function, reducing the weight of easy-to-classify samples and increasing the weight of difficult-to-classify samples.

[0017] As a further improvement of the present invention, in step 3.3, the global average pooling is specifically as follows:

[0018]

[0019] Where: F is the input feature map, H and W are the height and width of the feature map respectively, and c is the channel index;

[0020] The global maximum pooling is as follows:

[0021]

[0022] The ReLU activation function is as follows:

[0023] ReLU(x)=max(0,x)

[0024] The Sigmoid function is as follows:

[0025]

[0026] Where: e is the base of natural logarithm x is the input value;

[0027] In step 3.4, the focal loss function is as follows:

[0028] FocalLoss(p t )=-α(1-p t ) γ log(p t )

[0029] Where: p t is the model's predicted probability for each category. For positive samples, p t is the confidence of the model prediction as positive class; for negative samples, p t is the confidence of the model prediction as negative class; α t is the coefficient that balances the weights of positive and negative samples; γ is the focusing parameter that adjusts the weights of easy-to-classify samples and difficult-to-classify samples.

[0030] As a further improvement of the present invention, in step 4, the model is optimized by adjusting the data enhancement, attention mechanism, and loss function parameters in the improved algorithm to obtain a trained and optimized DCN-YOLO model, which is then deployed in practical applications to perform treeline fault identification tasks.

[0031] As a further improvement of the present invention, in step 5, the alarm system outputs "flame" and "branch" respectively when the trained and optimized DCN-YOLO model recognizes flames and branches, and uses "flame" and "branch" as input signals to trigger the alarm and enter the alarm system, thereby completing the early warning of tree line failure.

[0032] As a further improvement of the present invention, the early warning for tree line faults includes a primary warning and a secondary alarm. When the trained and optimized DCN-YOLO model only recognizes branches and outputs "branch", the primary warning is triggered. When the trained and optimized DCN-YOLO model recognizes flames and outputs "flame", the secondary alarm is triggered.

[0033] The beneficial effects of the present invention are:

[0034] 1. This invention simulates real treeline discharges using a treeline discharge experimental platform, captures videos, and exports images frame by frame. Labeling the images using labelimg creates an initial dataset. Roboflow acquires flame, cable, and branch datasets to expand the initial data into an enhanced dataset, which is then divided into training, validation, and test sets. Data augmentation, channel attention mechanism improvements, and Focal Loss sample weight balancing improvements are performed based on the DCN-YOLO model. The enhanced dataset is then fed into the improved DCN-YOLO model for training and validation, obtaining the highest mAP50 and mAP50-90 values ​​to develop a treeline fault monitoring and early warning model. This treeline fault monitoring and early warning model is then connected to a transmission tower monitoring system. The alarm system is triggered when branches, flames, or either of these are detected on the transmission line. By performing image recognition and early warning of treeline faults, this invention achieves accurate early warning of treeline fires, thereby reducing the probability of fires caused by treeline faults.

[0035] 2. The present invention uses an enhanced dataset containing a large number of labeled real tree-line discharge images as input into the DCN-YOLO model for training, and uses the improved DCN-YOLO model with the highest mAP50 and mAP50-90 values ​​as the tree-line fault monitoring and early warning model. This can quickly and accurately identify whether there are tree-line faults on the transmission line, realize real-time monitoring of the transmission line, and promptly discover potential tree-line fault hazards. Through the alarm system, it can more quickly and clearly confirm whether there are tree-line discharges on the transmission line, improve the integrated automation of fault identification and early warning, assist personnel in timely and effective prevention of the occurrence and expansion of tree-line faults, and improve the power supply reliability and safety of the transmission line. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Flowchart of an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0038] Example

[0039] like Figure 1 As shown, a tree line fault identification and warning method based on the DCN-YOLO model includes the following steps:

[0040] S1: Use the treeline discharge experimental platform to simulate real treeline discharges, shoot videos and export images by frame, and use labelimg to label the images to obtain the initial dataset;

[0041] The method for obtaining the initial data set is:

[0042] Image data from treeline faults is captured using image acquisition equipment installed around the transmission lines. This equipment includes mobile cameras and visual monitoring equipment installed at fixed nodes on the towers. The initial dataset is labeled with flames, cables, and branches.

[0043] S2: Use roboflow to obtain flame, cable, and tree branch datasets to expand the initial data to form an enhanced dataset, which is then divided into training, validation, and test sets.

[0044] The training set, validation set, and test set refer to dividing the enhanced dataset into a train training set, a val validation set, and a test test set in a ratio of 8:1:1.

[0045] S3: Data enhancement, channel attention mechanism improvement, and Focal Loss sample weight balance improvement based on the DCN-YOLO model;

[0046] The improvement based on the DCN-YOLO model includes the following steps:

[0047] S31: Build the DCN-YOLO model, which includes the backbone network responsible for extracting features from the input image, the head network responsible for making decisions on the target detection model and generating the final detection results, and the neck network responsible for feature fusion and enhancement. The deformable convolution DCNv3 is introduced to improve the performance of handling complex deformable targets.

[0048] S32: Randomly adjusts the color attributes (brightness, contrast, saturation, and hue) of the dataset in S2 to enhance the model's ability to recognize objects under varying lighting and color conditions. Images are randomly rotated by up to 15 degrees to simulate different possible orientations of objects in the image. A random region is cropped from the original dataset image and resized to 224x224 to help the model learn objects of varying sizes and proportions. Gaussian blur is applied to the dataset images to simulate images at different focal lengths, enhancing the model's invariance to details.

[0049] S33: A channel attention mechanism is integrated into the DCN-YOLO model. Global information of the feature map is obtained through global average pooling and global maximum pooling respectively. The weight relationship between channels is then learned through two 1×1 convolutional fully connected layers and ReLU activation function. Finally, the Sgmoid function is used to compress the output to the [0, 1] interval as the channel attention weight.

[0050] In S33, the attention module performs global average pooling, as shown in the following formula:

[0051]

[0052] Where: F is the input feature map, H and W are the height and width of the feature map respectively, and c is the channel index.

[0053] The global maximum pooling is shown in the following formula:

[0054]

[0055] Where: F is the input feature map, H and W are the height and width of the feature map respectively, and c is the channel index.

[0056] The ReLU activation function is shown in the following formula:

[0057] ReLU(x)=max(0,x)

[0058] The formula means that the ReLU function sets all negative values ​​to 0, while all non-negative values ​​remain unchanged.

[0059] The Sigmoid function is shown in the following formula:

[0060]

[0061] Where: e is the base of the natural logarithm, approximately equal to 2.71828, and x is the input value.

[0062] S34: Improve the model through the focal loss function, reducing the weight of easy-to-classify samples and increasing the weight of difficult-to-classify samples.

[0063] In S34, the focal loss function is expressed as follows:

[0064] Focal Loss(pt)=-α(1-pt)γlog(pt)

[0065] Where pt is the model's predicted probability for each class. For positive samples, pt is the confidence level of the model's prediction of the positive class; for negative samples, pt is the confidence level of the model's prediction of the negative class. αt is a coefficient that balances the weights of positive and negative samples. Different αt values ​​can be set for positive and negative samples, respectively. Typically, a smaller αt value is used for positive samples to address class imbalance. γ is a focusing parameter that adjusts the weights of easy-to-classify and hard-to-classify samples. It is usually a number greater than 0, reducing the loss contribution of easy-to-classify samples and increasing the loss contribution of hard-to-classify samples.

[0066] S4: The enhanced dataset is fed into the improved DCN-YOLO model for training and validation, obtaining the highest mAP50 and mAP50-90 values ​​and obtaining a treeline fault monitoring and early warning model.

[0067] In S4, by adjusting the data enhancement, attention mechanism, and loss function parameters in the improved algorithm to achieve the ideal model state, the trained and optimized DCN-YOLO model was obtained and deployed in practical applications to perform treeline fault identification tasks.

[0068] S5: The treeline fault monitoring and warning model is integrated into the transmission tower monitoring system. If a tree branch, a flame, or both are detected on the transmission line, an alarm system is triggered. By performing image recognition and early warning of treeline faults, this invention provides accurate early warning of treeline fires, thereby reducing the probability of fires caused by treeline faults.

[0069] In S5, the alarm system is triggered when a tree branch, a flame, or either of these is detected on the power line. The alarm system outputs "flame" and "branch" when the model recognizes flame and branch, respectively. Both "flame" and "branch" serve as input signals to the alarm system, providing early warning of tree line faults. The early warning is divided into primary and secondary warnings. A primary warning is triggered when the model only recognizes a tree branch and outputs "branch." A secondary warning is triggered when the model recognizes a flame and outputs "flame."

[0070] In this example, we first captured images of real treeline faults using image acquisition equipment installed around power lines. Labelimg was then used to label these images to generate an initial dataset. The roboflow public dataset was then used to augment the initial dataset with three categories: flames, cables, and branches. This enhanced dataset was then partitioned into a training set, a validation set, and a test set in an 8:1:1 ratio. A DCN-YOLO model was constructed, consisting of a backbone network responsible for extracting features from the input image; a head network responsible for making decisions and generating the final detection results; and a neck network responsible for feature fusion and enhancement. Deformable convolutional layer DCNv3 was introduced to improve performance in handling complex deformable targets. The color properties (brightness, contrast, saturation, and hue) of the augmented dataset are randomly adjusted to enhance the model's ability to recognize objects under varying lighting and color conditions. Images are randomly rotated up to 15 degrees to simulate different possible orientations of objects in the image. A region is randomly cropped from the original dataset image and resized to 224x224 to help the model learn objects of varying sizes and proportions. Gaussian blur is applied to the dataset images to simulate images at different focal lengths, enhancing the model's invariance to details. A channel-wise attention mechanism is integrated into the Yolov8 model. Global information from the feature maps is captured through global average pooling and global max pooling. Weights between channels are learned through two 1×1 convolutional fully connected layers and ReLU activation functions. Finally, the output is compressed to the range [0, 1] using the Sgmoid function, which serves as the channel attention weights. The model is improved using the focal loss function, which reduces the weight of easily classified samples and increases the weight of difficult samples. The enhanced dataset was fed into the improved DCN-YOLO model for training. By adjusting the data augmentation, attention mechanism, and loss function parameters in the improved algorithm to achieve the ideal model state, the resulting trained and optimized DCN-YOLO model was deployed in a real-world application for treeline fault identification. When the optimized model detects the presence of a branch, a flame, or either on a transmission line, an alarm system is triggered. The alarm system outputs "flame" and "branch" when the model detects a flame or a branch, respectively. Both "flame" and "branch" serve as triggering signals and enter the alarm system, providing early warning of treeline faults.

[0071] This example has the following advantages: the present invention uses an enhanced dataset containing a large number of real tree-line discharge images for labeling to input the DCN-YOLO model for training, and uses the improved DCN-YOLO model with the highest mAP50 and mAP50-90 values ​​as the tree-line fault monitoring and early warning model. It can quickly and accurately identify whether there is a tree-line fault on the transmission line, realize real-time monitoring of the transmission line, and promptly discover potential tree-line fault hazards; through the alarm system, it can more quickly and clearly confirm whether there is a tree-line discharge in the transmission line, improve the integrated automation of fault identification and early warning, assist staff to timely and effectively avoid the occurrence and expansion of tree-line faults, and improve the power supply reliability and safety of the transmission line.

[0072] The above-described embodiments merely represent specific implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A tree line fault identification and early warning method based on the DCN-YOLO model, characterized in that: The following steps are involved: Step 1: Use the treeline discharge experimental platform to simulate real treeline discharges, shoot videos, export images frame by frame, and label the images to obtain the initial dataset; Step 2: Obtain the flame, cable, and tree branch datasets to expand the initial dataset to form an enhanced dataset, and divide it into a training set, a validation set, and a test set; Step 3: Based on the DCN-YOLO model, perform data enhancement, improve the channel attention mechanism, and improve the loss function sample weight balance; The step 3 specifically includes the following steps: Step 3.

1. Build the DCN-YOLO model, which includes the backbone network for extracting features from the input image, the head network for making decisions and generating the final detection results, and the neck network for feature fusion and enhancement. The deformable convolutional layer DCNv3 is introduced to improve the performance of handling complex deformable targets. Step 3.2: Randomly adjust the color attributes of the dataset to enhance the DCN-YOLO model's ability to recognize targets under different lighting and color conditions. Also, randomly rotate the image by up to 15 degrees to simulate different possible orientations of the target in the image. Randomly crop a region from the original dataset image and resize it to 224x224 for the DCN-YOLO model to learn targets of different sizes and proportions. Apply Gaussian blur to the dataset image to simulate images at different focal lengths and enhance the DCN-YOLO model's invariance to details. Step 3.3: Integrate a channel attention mechanism into the DCN-YOLO model. Use global average pooling and global maximum pooling to obtain global information of the feature map. Then, use two 1×1 convolutional fully connected layers and ReLU activation functions to learn the weight relationship between channels. Finally, use the Sgmoid function to compress the output to the [0, 1] interval as the channel attention weight. Step 3.4: Improve the DCN-YOLO model by using the focal loss function to reduce the weight of easy-to-classify samples and increase the weight of difficult-to-classify samples. Step 4: Input the enhanced dataset into the improved model for training and verification, obtain the highest evaluation index mAP50 and mAP50-90 values, and obtain the tree line fault monitoring and early warning model; Step 5: Connect the tree-line fault monitoring and early warning model to the transmission tower monitoring system. When branches, flames, or either of the two are detected on the transmission line, the alarm system will be triggered.

2. The tree line fault identification and early warning method based on the DCN-YOLO model according to claim 1 is characterized in that: In step 1, image data of a tree line fault occurring is obtained by using image acquisition equipment installed around the transmission line, wherein the image acquisition equipment includes a movable camera and a visual monitoring device installed at a fixed node of a tower; The initial data set refers to the data set of flame, cable, and branch labels after labeling.

3. The tree line fault identification and early warning method based on the DCN-YOLO model according to claim 1 is characterized in that: In step 3.3, global average pooling is performed as follows: Where: F is the input feature map, H and W are the height and width of the feature map respectively, and c is the channel index; The global maximum pooling is as follows: The ReLU activation function is as follows: ReLU(x)=max(0,x) The Sigmoid function is as follows: Where: e is the base of natural logarithm x is the input value; In step 3.4, the focal loss function is as follows: FocalLoss(p t )=-α(1-p t ) γ log (p t ) Where: p t is the model's predicted probability for each category. For positive samples, p t is the confidence of the model prediction as positive class; for negative samples, p t is the confidence of the model prediction as negative class; α t is the coefficient that balances the weights of positive and negative samples; γ is a focusing parameter that adjusts the weights of easy-to-classify samples and difficult-to-classify samples.

4. The tree line fault identification and early warning method based on the DCN-YOLO model according to claim 1 is characterized in that: In step 4, the data enhancement, attention mechanism, and loss function parameters in the improved algorithm are adjusted to achieve the ideal model state, and the trained and optimized DCN-YOLO model is obtained and deployed in practical applications to perform treeline fault identification tasks.

5. The tree line fault identification and early warning method based on the DCN-YOLO model according to claim 4 is characterized in that: In step 5, the alarm system outputs "flame" and "branch" respectively when the trained and optimized DCN-YOLO model recognizes flames and branches, and uses "flame" and "branch" as input signals to trigger the alarm, thereby providing early warning of tree line failures.

6. The tree line fault identification and early warning method based on the DCN-YOLO model according to claim 5 is characterized in that: The early warning for tree line failure includes primary warning and secondary alarm. When the trained and optimized DCN-YOLO model only recognizes branches and outputs "branch", the primary warning is triggered. When the trained and optimized DCN-YOLO model recognizes flames and outputs "flame", the secondary alarm is triggered.

Images