Power transmission line bird identification method based on GMM foreground detection and neural network
By using a combination of Gaussian hybrid model and convolutional neural network on the transmission line, bird behavior and bird species are recognized, solving the problem of identifying bird-related faults in the power grid, and improving the operating safety of the power grid.
Patent Information
- Application Number
- CN202411165463.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-11-29
AI Technical Summary
The existing power transmission line operation and maintenance personnel lack bird identification methods and are unable to detect bird-related faults and identify bird species in a timely manner, making it difficult to achieve precise prevention and control of bird-related faults in the power grid.
A bird recognition model is established based on Gaussian hybrid model (GMM) for prospect detection and convolutional neural network (CNN), and the bird behavior is detected and species are identified through image recognition methods, and the association between bird activities and bird-related failures is established.
Real-time detection and bird species identification of bird behavior in the power grid are realized, effectively preventing and preventing bird-related failures, and improving the operating safety and stability of the power grid.
Smart Images

Figure CN119048907A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of power system operation safety protection, and in particular relates to a transmission line bird recognition method based on GMM foreground detection and neural network. Background Art
[0002] With the construction of power grid lines and the development of ecological environment, the contradiction between the activity habits of birds and the operation of lines has been increasing, and birds have frequently caused power system operation failures. Bird-related failures in power systems may include bird nesting, bird droppings, bird body short circuits and other types of failures, which have a great impact on the safe and stable operation of power systems.
[0003] Bird-related faults in transmission circuits and towers are random and sudden, and are closely related to local bird species and bird habits. When a bird-related fault occurs in the system, operation and maintenance personnel usually find it difficult to discover it in time and cannot determine what kind of bird behavior caused the fault, which is not conducive to fault analysis and prevention measures by power system inspectors. Therefore, it is necessary to conduct real-time image monitoring of power towers and transmission lines, effectively detect bird behavior and identify bird species, establish a correlation between bird activities and bird-related faults, and realize accurate prevention and control of bird-related faults in the power grid. Summary of the invention
[0004] The technical problem to be solved by the present invention is that the existing transmission line operation and maintenance personnel lack a method for bird identification, and are unable to timely discover bird-related faults, effectively detect bird behaviors and identify bird species, and thus cannot effectively achieve accurate prevention of bird-related faults in power grids. The present invention provides a transmission line bird identification method based on GMM foreground detection and neural network, comprising the following steps:
[0005] S1, collect the on-site monitoring video of the transmission line to obtain historical image data;
[0006] S2, establish a foreground detection model based on Gaussian mixture model (GMM);
[0007] S3, inputting the historical image data collected in S1 into the GMM-based foreground detection model to obtain foreground images, and dividing the obtained foreground image data into a test set and a training set;
[0008] S4, using a linear interpolation method to uniformly adjust the pixel size of the foreground image obtained in S3;
[0009] S5, establishing a bird recognition model based on a convolutional neural network (CNN); inputting the training set adjusted by S4 into the bird recognition model based on the CNN to complete the training of the bird recognition model based on the CNN;
[0010] S6, Model functional verification.
[0011] Furthermore, S2 specifically includes constructing an image foreground detection model based on GMM and initializing and updating parameters of the image foreground detection model based on GMM.
[0012] Furthermore, in S2, the monitoring video collected at the transmission line site is extracted according to the time dimension.
[0013] Furthermore, in S2, by fitting the probability distribution of the background and foreground pixels of the historical image data, the GMM foreground detection model can distinguish the foreground and background in the image;
[0014] Furthermore, S5 specifically includes CNN-based image classification model building and model parameter update
[0015] Furthermore, in S5, the trained CNN-based bird recognition model is obtained by optimizing parameters through forward propagation and back propagation using the Adam method;
[0016] Furthermore, the model function is verified: the test set image data is input into the GMM-based foreground detection model and the trained CNN-based bird recognition model. Through actual case analysis, the bird species recognition function of the GMM-based foreground detection model and the CNN-based bird recognition model is verified.
[0017] Through the above design scheme, the present invention can bring the following beneficial effects:
[0018] The present invention establishes a foreground detection model based on a Gaussian mixture model (GMM) and a bird recognition model based on a convolutional neural network. Through image recognition methods, the behavior of birds during the operation of the power system is detected, the types of bird species are effectively identified, and the bird activities are associated with bird-related faults, thereby achieving accurate prevention and control of bird-related faults in the power grid. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:
[0020] Figure 1 is the flow chart of the proposed method;
[0021] Figure 2 It is the GMM foreground detection map;
[0022] Figure 3 It is the CNN input image recognition result. DETAILED DESCRIPTION
[0023] The present invention provides a method for identifying birds on power transmission lines based on GMM foreground detection and neural network, comprising the following steps:
[0024] S1, collect on-site monitoring videos of the transmission line to obtain historical image data; install a monitoring camera on one side of the transmission line tower, and the monitoring camera is used to collect and transmit video information of birds staying near the tower;
[0025] S2, establishing a foreground detection model based on a Gaussian mixture model (GMM); the foreground detection model based on the Gaussian mixture model (GMM) fits the probability distribution of the background and foreground pixels of the historical image data, so that the GMM foreground detection model can distinguish the foreground and background in the image, and can classify the pixels in the image of the input power transmission line, thereby distinguishing the foreground and the background, and thus extracting the foreground image;
[0026] S2 specifically includes step S21: constructing an image foreground detection model based on GMM and step S22: initializing and updating parameters of the image foreground detection model based on GMM. train ={x train ,1,…,x train,T}, traverse all images, T is the number of images, calculate the GMM model through formulas (2)-(9), update the model parameters, and obtain the foreground of the input image.
[0027] S21, image foreground detection model construction based on GMM:
[0028] Image foreground detection through GMM is a process of modeling the image background in a specific scene. First, set x to be a random process, that is, the value of the pixel in the time dimension at a given coordinate. A Gaussian mixture model is established for the probability density function of x, which consists of the weighted composition of K Gaussian distributions. If p(x) is a probability density function composed of K Gaussian distributions, then the GMM model corresponding to x is:
[0029]
[0030] In the formula, w k is the weight corresponding to the kth Gaussian component, μ k , σ k are the mean and standard deviation corresponding to the kth Gaussian component, N(x; μ k ,σ k ) indicates that for the kth Gaussian distribution in the GMM model of x, the mean is μ k , the covariance matrix is I is the identity matrix.
[0031] S22. Parameter initialization and update of GMM-based image foreground detection model:
[0032] In order to obtain a robust representation of image background modeling, the model parameters (wk , μ k , σ k ) is updated and changes with time. First, the model parameters are initialized according to formula (2):
[0033]
[0034] In the formula, w k0 , μ k0 , σ k0 For initialization parameters.
[0035] For x t , Formula (3) is the judgment condition for updating the parameters of the j-th Gaussian component, and the calculation formula is:
[0036] ||x t -μ j || 2 / σ j <τ,j∈[1..K] (3)
[0037] If component j meets the discrimination condition, update the GMM model parameters, and the calculation formula is:
[0038]
[0039] μ k (t)=(1-β)μ k (t-1)+βx t (5)
[0040]
[0041] When the judgment condition of formula (3) is met, Then determine x t The pixel is the background. If the condition is not met This pixel is the foreground. After traversing all the pixels in the image, the foreground detection function in the test image can be realized. If all components j do not meet the conditions, all parameters of the component m corresponding to the lowest weight among all components are reinitialized, that is:
[0042]
[0043] Where α represents the learning rate, which is usually set to a constant that determines the component weight w k The size of the change. β is the parameter for updating the mean and variance step length, and the calculation formula is:
[0044] β=αN(x t ;μ k ,σ k )(8)
[0045] In each iteration of updating parameters, the weight w k Normalize the weights so that the sum of all weights is 1. The calculation formula is:
[0046]
[0047] Furthermore, the image dataset used is extracted according to the time dimension based on the transmission line monitoring video collected on-site. In the image set used, the proportion of background pixels far exceeds the proportion of foreground pixels, which is conducive to the application effect of the GMM-based foreground detection model.
[0048] S3, input the historical image data collected in S1 into the GMM-based foreground detection model to obtain foreground images, divide the obtained foreground image data into a test set and a training set in a ratio of 8:2 and assign labels;
[0049] S4, using the linear interpolation method to uniformly adjust the pixel size of the training set foreground image data obtained in S3; the GMM model detection result is the foreground feature map in the original image. Since the foreground pixel size is not uniform, it is not conducive to the subsequent training and classification of the CNN model. The pixel size of the feature map is adjusted by equations (10)-(13) to obtain a batch of feature maps with uniform pixel size; the extracted image foreground is pixel-adjusted so that the extracted foreground image has the same pixel size, which is convenient for the subsequent CNN classification model training and classification.
[0050] For a foreground image x i , resize it to a pixel size of c 1 ×c 2 The image is represented as:
[0051] r i (a 0 ,e 0 )(10)
[0052] In the formula, r i (a 0 ,e 0 ) is the pixel value with coordinates a0, e0 in the adjusted image. If the pixel point to be calculated is between the image pixels (a', e'), (a', e'+1), (a'+1, e'), (a'+1, e'+1), then r i (a 0 ,e 0 ) is calculated as:
[0053]
[0054]
[0055]
[0056] For all x i , i=1,2,…,n to adjust the pixels, then the image size of the adjusted foreground image set X' is c 1 ×c 2 .
[0057] S5, establish a bird recognition model based on convolutional neural network (CNN). S5 specifically includes S51: CNN-based image classification model modeling and S52: model parameter update. In S5, the trained CNN-based bird recognition model is obtained by forward propagation and back propagation and by optimizing parameters through the Adam method. After the construction and training of the CNN classification model, the CNN model structure is determined, the CNN hyperparameters are set, and the feature maps with labels and uniform sizes are input into the CNN for training. The network is forward propagated through formulas (14)-(18), the model back propagation gradient is calculated through formulas (19)-(23), and the model parameters are updated through formulas (24)-(29). When the maximum number of iterations is reached, the model training is completed and the CNN-based bird recognition model is obtained. The trained CNN-based bird recognition model can realize the bird species classification function after the bird image is input.
[0058] S51. Modeling of image classification model based on CNN:
[0059] The structure of the image classification model based on CNN consists of an input layer, a convolutional layer, an activation function, a pooling layer, a fully connected layer, and an output layer. The structure of the image recognition model based on the convolutional neural network is determined. The specific calculation process is as follows:
[0060] (1) Input layer. The input image is represented as a three-dimensional matrix X k(H×E×C) , H is the height of the image, E is the width of the image, and C is the number of RGB channels.
[0061] (2) Convolution layer. A finite number of convolution kernels are used to perform convolution operations on the input original image data. Each convolution kernel can extract different features. The convolution calculation formula is:
[0062]
[0063] Where X k is the input image, g is the convolution kernel, (i, j) is the output position, and the convolution result is the feature map extracted by the convolution layer.
[0064] (3) Activation function
[0065] CNN often uses ReLU (Rectified Linear Unit) function as the activation function, and its calculation formula is:
[0066] f relu (x)=max(0,x) (15)
[0067] x is the input of the ReLU function. For all negative values in the input x, f relu is 0.
[0068] (4) Pooling layer
[0069] CNN often uses maximum pooling or average pooling methods to reduce the size of feature maps, while reducing the data dimension and retaining the main features to the greatest extent.
[0070] The calculation formula of the maximum pooling method is:
[0071] y(i,j)=max (m,n)∈pool x(m,n) (16)
[0072] Where x is the input feature map, y is the feature map after pooling, and pool is the pooling window.
[0073] (5) Fully connected layer
[0074] The pooled feature map data is flattened and forward propagated through a series of fully connected layers. The calculation formula for the output of the fully connected layer is:
[0075] y=f σ (Wx+b) (17)
[0076] Where W is the weight matrix for input x, x is the input, b is the bias, and f σ It is the activation function. ReLU function or Sigmoid function is usually used as the activation function of the fully connected layer.
[0077] (6) Output layer
[0078] For multi-classification tasks, the Softmax function is often used as the output of the output layer, and the calculation formula is:
[0079]
[0080] In the formula, z i It is the i-th component in the output vector, and the output vector is converted into probability through the Softmax function.
[0081] Through the above process and the calculation formulas of each layer for forward propagation, CNN can extract high-order features from the input image and classify it through the output layer. During the back-propagation process, the weights and bias parameters of each layer are updated to minimize the classification error.
[0082] S52, model parameter update:
[0083] In the parameter update process of model training iteration, it is necessary to determine the loss function, back propagation gradient calculation method and parameter optimization method. The cross entropy loss is used as the loss function of the CNN model, and the calculation formula is:
[0084]
[0085] In the formula, y i is the actual image classification label, is the predicted classification probability.
[0086] (1) Back propagation calculation, calculate the gradient of the loss function for each parameter through the chain rule.
[0087] The output layer gradient is:
[0088]
[0089] The fully connected gradient is:
[0090]
[0091] The convolution kernel gradient calculation formula is:
[0092]
[0093] Gradient of the ReLU function:
[0094]
[0095] f 1,x>0 is the indicator function. When x>0, f 1,x>0 =1, otherwise it is 0.
[0096] (2) Parameter update method: Based on the gradient calculated in step (1), the parameters are updated using the Adam optimizer. For each iteration t, the parameter calculation process is as follows:
[0097] 1) Calculate the gradient g of the gradient loss function L with respect to the parameter θ t
[0098]
[0099] 2) Update the first-order moment estimation parameter m t and the second-order moment estimate parameter v t
[0100] m t =β 1 m t-1 +(1-β 1 ) t (25)
[0101]
[0102] 3) Calculate the deviation m of the first-order moment and second-order moment estimates t '、v t '
[0103]
[0104]
[0105] 4) Update parameter θ
[0106]
[0107] In the formula, θ t is the updated parameter at iteration number t. 1 , β 2 are the decay rates of the first-order moment estimate and the second-order moment estimate, respectively. ∈ is a very small constant, usually 10 -8 , the learning rate η is used to control the step size of parameter update.
[0108] The Adam parameter optimization method can adaptively adjust the learning rate, making the parameter update process more reliable and efficient. It is widely used in practical applications.
[0109] S7, test image recognition:
[0110] The test set image data is input into the GMM-based foreground detection model and the trained CNN-based bird recognition model to verify the model functions. Through actual example analysis, the foreground detection function of the GMM model and the bird species recognition function of the CNN-based bird recognition model are verified.
[0111] In this embodiment, the test set image is input into the foreground detection model based on GMM, and the obtained foreground feature map is shown in the attached figure. Figure 2 , after adjusting the pixel size of the foreground feature map in step S4, the following is obtained: Figure 3 The foreground image will be Figure 3 The foreground image is input into the CNN-based bird recognition model, and the corresponding category obtained is magpie, which verifies the effectiveness of the model.
Claims
1. A method for identifying birds on power transmission lines based on GMM foreground detection and neural network, characterized in that: The following steps are involved: S1, collect the on-site monitoring video of the transmission line to obtain historical image data; S2, establish a foreground detection model based on Gaussian mixture model (GMM); S3, inputting the historical image data collected in S1 into the GMM-based foreground detection model to obtain foreground images, and dividing the extracted foreground image data into a training set and a test set; S4, using a linear interpolation method to uniformly adjust the pixel size of the foreground image obtained in S3; S5, establishing a bird recognition model based on a convolutional neural network (CNN); inputting the training set adjusted by S4 into the bird recognition model based on the CNN to complete the training of the bird recognition model based on the CNN; S6, Model functional verification.
2. The method for identifying birds on power transmission lines based on GMM foreground detection and neural network according to claim 1, characterized in that: S2 specifically includes the construction of a GMM-based image foreground detection model and the initialization and updating of parameters of the GMM-based image foreground detection model.
3. The method for identifying birds on power transmission lines based on GMM foreground detection and neural network according to claim 2, characterized in that: In S2, the monitoring video collected at the transmission line site is extracted according to the time dimension.
4. The method for identifying birds on power transmission lines based on GMM foreground detection and neural network according to claim 2, characterized in that: In S2, by fitting the probability distribution of the background and foreground pixels of the historical image data, the GMM foreground detection model can distinguish the foreground and background in the image.
5. The method for identifying birds on power transmission lines based on GMM foreground detection and neural network according to claim 1, characterized in that: S5 specifically includes CNN-based image classification model building and model parameter updating.
6. The method for identifying birds on power transmission lines based on GMM foreground detection and neural network according to claim 5, characterized in that: In S5, the trained CNN-based bird recognition model is obtained by forward propagation and back propagation and by optimizing parameters using the Adam method.
7. The method for identifying birds on power transmission lines based on GMM foreground detection and neural network according to claim 1, characterized in that: In S6, model function verification: the test set image data is input into the GMM-based foreground detection model and the trained CNN-based bird recognition model. Through actual case analysis, the bird species recognition function of the GMM-based foreground detection model and the CNN-based bird recognition model is verified.
Citation Information
Patent Citations
A CNN and MOG-based transformer substation foreign matter invasion detection system
CN109785361A
Power transmission line multi-target detection method
CN112184692A
Weak and small airspace target detection method based on super-resolution feature enhancement
CN113223059A
Power transmission line nest droop detection method, system and device and medium
CN117455899A
