Transformer discharge carbon mark detection method based on VGG-16 pruned model

Abstract

This invention relates to a transformer discharge carbon mark detection method based on the VGG-16 pruning model. A microrobot captures the original internal image of the transformer along a planned path within the oil-immersed transformer tank. After image and data enhancement, a VGG-16 network model is established. Comparative experiments are conducted on VGG-16 network pruning using the L1-norm pruning algorithm and the FPGM pruning algorithm. Based on the experimental results, a pruning algorithm model suitable for transformer internal carbon mark detection is ultimately determined. The transformer is then tested for carbon mark. While meeting the accuracy requirements for actual transformer carbon mark detection, this method also meets the requirements of industrial scenarios for low hardware resource usage and high speed.

Description

Technical Field

[0001] The present invention belongs to the technical field of power transmission equipment detection, and in particular to a transformer discharge carbon mark detection method based on a VGG-16 pruning model. Background Art

[0002] Large power transformers typically use oil-impregnated cardboard as internal insulation. During operation, partial discharge (PD) can cause carbon marks on the insulation, reducing the effective insulation distance and seriously threatening the transformer's safety. The transformer's metal casing is enclosed, making it difficult for conventional PD signal monitoring devices to detect carbon marks inside the oil tank. To address this issue, it is essential to capture surface images of components such as the transformer's internal windings and cardboard to examine the amount of carbon marks and visually determine the location of insulation faults and the degree of insulation degradation. Identifying carbon marks within these captured images of the transformer's interior provides effective data support for determining the transformer's insulation condition. Summary of the Invention

[0003] The purpose of the present invention is to overcome the shortcomings of the existing technology and propose a transformer discharge carbon mark detection method based on the VGG-16 pruning model. The method uses machine vision and deep learning to detect and identify the internal discharge carbon marks of the transformer. The L1-norm algorithm and the FPGM algorithm are used to conduct network pruning experiments on VGG-16 respectively, and a model is selected from them to perform carbon mark detection on the transformer. On the basis of meeting the actual accuracy requirements of transformer carbon mark detection, the present invention meets the requirements of industrial scenarios for less hardware resource occupation and fast operation speed.

[0004] The present invention solves the technical problem by adopting the following technical solutions:

[0005] The transformer discharge carbon mark detection method based on the VGG-16 pruned model includes the following steps:

[0006] Step 1: The micro robot takes the original internal image of the transformer along the planned path inside the oil tank of the oil-immersed transformer and transmits it back to the control platform;

[0007] Step 2: The control platform uses image enhancement and data enhancement to process the dataset constructed by the image acquired in step 1, performs image enhancement on the original dataset, and then expands the quantity of the dataset after image enhancement to obtain the dataset used in the experiment;

[0008] Step 3. Establish a VGG-16 network model and conduct pruning comparison experiments of the L1-norm pruning algorithm and the FPGM pruning algorithm under different pruning rate settings. Based on the experimental results, the pruning algorithm model suitable for carbon mark detection inside the transformer is finally determined, and carbon mark detection is performed inside the transformer.

[0009] Moreover, the specific implementation method of the image enhancement in step 2 is: enhancing the image by improving the histogram equalization algorithm, and the specific implementation method of the data enhancement in step 2 is: expanding the image enhanced data set by using four methods: Gaussian blur processing, horizontal and vertical flipping, scale scaling, and horizontal and vertical translation, and performing data enhancement of image fusion according to the Mosaic algorithm.

[0010] Moreover, the improved histogram equalization algorithm for image enhancement includes the following steps:

[0011] Step 2.1, divide the image obtained in step 1 into 8 parts horizontally and 4 parts vertically;

[0012] Step 2.2, calculate the histogram of each sub-block, and determine the dynamic range and emphasis factor of each histogram;

[0013] New grayscale value dynamic range r i The calculation method is:

[0014]

[0015]

[0016] c i =S max -S min

[0017] Among them, c i is the gray level range of each sub-block histogram, w i is the emphasis factor of each sub-block histogram, n i The total number of pixels in each sub-block, m i is the number of pixel types in each sub-block, W k is the emphasis factor of the kth gray level, t is the total gray value level, L is the number of divided sub-blocks, S max The maximum grayscale value in each sub-block histogram; S min The minimum grayscale value in each sub-block histogram;

[0018] Step 2.3, perform independent grayscale equalization on each sub-histogram in step 2.1;

[0019] Perform histogram equalization on each sub-block, and its mapping function is:

[0020]

[0021] Among them, S i is the grayscale value of the mapped image; n is the total number of pixels in the sub-block; n k is the number of pixels of gray level k in the sub-block.

[0022] Furthermore, the specific implementation method for data augmentation using the Mosaic algorithm for image fusion is as follows: randomly select one image from the group labeled with carbon marks and perform size reduction, randomly select three other images from the entire dataset and perform random cropping, and then stitch these four images into a single image with the classification label of "with carbon marks." This method generates several images labeled with carbon marks, enriching the image background and reducing the size of large carbon marks.

[0023] Moreover, the specific implementation method of step 3 is: establishing a VGG-16 network model, performing network pruning with different pruning rates based on the L1-norm pruning algorithm and the FPGM pruning algorithm for the experimental data set obtained in step 2, obtaining the optimal pruning algorithm through comparison, and performing carbon mark detection on the inside of the transformer through the optimal pruning algorithm.

[0024] Moreover, the specific calculation method of the L1-norm pruning algorithm in step 3 is:

[0025] The objective function of the L1 norm regularization term for:

[0026]

[0027]

[0028] Among them, J(w) is the original objective function; w is the weight matrix transferred between each layer, x i is the input vector; y i is the expected output; α is a hyperparameter that weighs the relative contribution of the regularization term to the objective function J, and its range is α∈[0,∞); m is the mth parameter.

[0029] Objective gradient function for:

[0030]

[0031] Where w* is the weight matrix caused by the regularization term; H is the Hessian matrix of J(w) at w*; the Hessian matrix is ​​a diagonal matrix, H = diag([H 1,1 ,...,H n,n ]), each element on the diagonal is a positive number; the quadratic approximation of the objective function is decomposed into:

[0032]

[0033]

[0034] Among them, H i,iis the i-th element on the diagonal of the Hessian matrix; w i is the function weight.

[0035] when When the original objective function J(w) of the convolution in the i-th dimension is The contribution of w is suppressed, and the L1 regularization term i Push to 0, w i The optimal value of is 0; when When , the L1 regularization term will not change w i Push to 0 and move on the original basis

[0036] After the L1-norm pruning algorithm sets the pruning rate, it stops pruning for each layer when the set value is reached, and fine-tunes the pruned model to compensate for potential temporary degradation.

[0037] Moreover, the specific method of the FPGM pruning algorithm in step 3 is as follows: updating the network weight parameters, setting the epoch value and pruning rate; starting from the first epoch, during the training process, the redundant convolution kernels are set to zero according to the geometric median and pruning rate of each layer; the zeroed convolution kernels are re-updated with parameters in the next epoch until the epoch execution completes the set value, and iterates repeatedly until the redundant filters tend to be sparse.

[0038] The advantages and positive effects of the present invention are:

[0039] 1. The present invention uses a microrobot to capture the original internal image of the transformer along a planned path inside the oil tank of an oil-immersed transformer. After image enhancement and data enhancement, a VGG-16 network model is established. A network pruning comparison experiment is conducted on VGG-16 using the L1-norm pruning algorithm and the FPGM pruning algorithm. Based on the experimental results, a pruning algorithm model suitable for carbon mark detection inside the transformer is finally determined, and carbon mark detection is performed on the transformer. On the basis of meeting the accuracy requirements of actual transformer carbon mark detection, the present invention meets the requirements of industrial scenarios for low hardware resource usage and high operation speed, and is suitable for deployment in actual embedded devices.

[0040] 2. To address the issues of low contrast and difficulty in extracting target features in the original images collected by the microrobot, this paper uses an improved histogram equalization algorithm to enhance the dataset. This effectively improves the image contrast and clarity, while preserving target details and suppressing noise, making image features more distinct and easier to extract. Furthermore, the Mosaic algorithm is used for data enhancement after image enhancement, increasing the number of samples and enriching the diversity of the dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flow chart of the present invention;

[0042] Figure 2 Schematic diagram of L1-norm algorithm pruning. DETAILED DESCRIPTION

[0043] The present invention is further described below in conjunction with the accompanying drawings.

[0044] Transformer discharge carbon mark detection method based on VGG-16 pruning model, such as Figure 1 As shown, the following steps are included:

[0045] Step 1: The micro robot takes the original internal image of the transformer along the planned path inside the oil tank of the oil-immersed transformer and transmits it back to the control platform.

[0046] The specific implementation method of this step is: use a micro-bionic fish to conduct autonomous patrols along the planned path inside the oil-immersed transformer tank. The bionic fish head is equipped with an LED light source and an industrial camera to capture raw data images of the transformer along the way and transmit them back to the control platform.

[0047] Step 2: The control platform uses image enhancement and data enhancement to process the dataset constructed by the image obtained in step 1, performs image enhancement on the original dataset, and then expands the quantity of the image-enhanced dataset to obtain the dataset used in the experiment.

[0048] The images acquired in step 1 have the following two characteristics: The image fill level varies depending on the shooting distance and angle. Some images taken at longer distances have low contrast and are generally darker, making feature extraction difficult. The data volume is relatively small, and the positive and negative samples are unbalanced, which can easily lead to model overfitting and poor generalization. Therefore, further processing is required to obtain a dataset that meets the requirements of the present invention. Therefore, the present invention processes the dataset using image enhancement using improved histogram equalization and data enhancement using image fusion based on the Mosaic concept.

[0049] The histogram equalization algorithm uses the cumulative distribution function of the probability of each gray level in the image as a transformation function to perform a mapping transformation on the original image, making the originally overly concentrated gray levels evenly distributed. Based on this, the present invention proposes an improved histogram equalization algorithm. The improved histogram equalization algorithm for image enhancement includes the following steps:

[0050] Step 2.1: Divide the image obtained in step 1 into 8 parts horizontally and 4 parts vertically, that is, the original image will be divided into 4×8=32 sub-blocks;

[0051] Step 2.2, calculate the histogram of each sub-block, and determine the dynamic range and emphasis factor of each histogram;

[0052] Calculate the grayscale range c of each sub-block histogram i :

[0053] c i =S max -S min

[0054] Among them, S max The maximum grayscale value in each sub-block histogram; S min The minimum grayscale value in each sub-block histogram;

[0055] According to the emphasis factor w of each sub-block histogram i , calculate the new gray value dynamic range r i

[0056]

[0057]

[0058] Among them, n i The total number of pixels in each sub-block, m i is the number of pixel types in each sub-block, W k is the emphasis factor of the kth gray level, t is the total gray value level, and L is the number of divided sub-blocks;

[0059] Step 2.3, perform independent grayscale equalization on each sub-histogram in step 2.1;

[0060] Perform histogram equalization on each sub-block, and its mapping function is:

[0061]

[0062] Among them, S i is the grayscale value of the mapped image; n is the total number of pixels in the sub-block; n k is the number of pixels of gray level k in the sub-block.

[0063] Image enhancement is to use homomorphic filtering, histogram equalization and the algorithm of this paper to enhance the original image respectively. Comparison shows that: the overall brightness of the original image is low, the dynamic range of grayscale values ​​is narrow, and the outline and color of the carbon mark are almost indistinguishable; the image processed by homomorphic filtering effectively improves the brightness of the original image, so that the carbon mark in the center of the image can be distinguished, but the overall grayscale value is high, and the carbon mark target is not obvious; the images processed by histogram equalization and the method of the present invention both effectively increase the dynamic range of the image grayscale, so that the image brightness and contrast are significantly improved, and the carbon mark area is significantly different from the surrounding background; but there is a problem of detail loss in the carbon mark area in the histogram equalization, and a problem of noise enhancement in the aperture area. This is because the histogram equalization algorithm will cause some low-frequency grayscales to be merged with the prominent grayscale, while the local and detailed processing of the image processed by the method of the present invention is more detailed, the brightness of the fill light area is uniform, and the details and features of the target area are better preserved. It can be seen that the improved histogram equalization algorithm adopted in the present invention not only effectively increases the dynamic range of the image grayscale, significantly improves the average grayscale and contrast of the image, but also effectively suppresses local noise, making the image features more obvious and easier to extract.

[0064] Data augmentation is performed on the enhanced dataset. This is achieved by expanding the dataset using four methods: Gaussian blurring, horizontal and vertical flipping, scaling, and horizontal and vertical translation. This brings the total dataset to 2,700 images. This increases the number of samples and prevents overfitting. To enhance sample richness, the present invention also uses the Mosaic algorithm for image fusion data augmentation.

[0065] The specific data augmentation method involves randomly selecting one image from the "carbon trace" group and reducing its size. Three other images from the entire dataset are randomly cropped and then concatenated into a single image with the "carbon trace" classification label. This method generates 1,600 images labeled "carbon trace," enriching the image context and reducing the size of large carbon traces. This better balances the scale distribution of the training data, enabling the model to effectively improve the detection accuracy of small objects.

[0066] Step 3. Establish a VGG-16 network model and conduct pruning comparison experiments of the L1-norm pruning algorithm and the FPGM pruning algorithm under different pruning rate settings. Based on the experimental results, the pruning algorithm model suitable for carbon mark detection inside the transformer is finally determined, and carbon mark detection is performed inside the transformer.

[0067] The established VGG-16 network model is as follows: the VGG-16 model takes a 224×224×3 image as input. It first extracts detailed features of the image through two 64-channel convolutional layers and a max pooling layer. It then extracts intermediate features through two 128-channel convolutional layers and a max pooling layer. Finally, three convolutional layers and a max pooling layer are connected to gradually extract classification features. The most distinctive feature of this model is that all convolutional kernels use small 3×3 kernels. Feature extraction is achieved through repeated deepening of convolutional and pooling layers. Small convolution kernels extract more delicate details, helping to improve the recognition rate of small objects. However, due to the increased number of layers, the number of parameters reaches 138.4MB. Recognizing a single image requires approximately 554MB of memory and 15 billion floating-point calculations.

[0068] In order to reduce the number of model parameters and ensure the real-time classification, the VGG-16 network model was pruned using the L1-norm pruning algorithm and the FPGM pruning algorithm.

[0069] The L1-norm algorithm ranks the importance of convolution kernels according to the L1 norm, removes convolution kernels that have little impact on model accuracy, and achieves coefficient regularization of the model by deleting the entire convolution kernel and its connected feature graph in the network. The specific calculation method of the L1-norm pruning algorithm is:

[0070] The L1 norm refers to the sum of the absolute values ​​of each element in the vector, also known as the "sparse regularization operator". The objective function of the L1 norm regularization term is introduced for:

[0071]

[0072]

[0073] Among them, J(w) is the original objective function; w is the weight matrix transferred between each layer, x i is the input vector; y i is the expected output; α is a hyperparameter that weighs the relative contribution of the regularization term to the objective function J, and its range is α∈[0,∞); m is the mth parameter.

[0074] Objective gradient function for:

[0075]

[0076] Where w* is the weight matrix caused by the regularization term; H is the Hessian matrix of J(w) at w*; in order to simplify the derivation steps without losing generality, the Hessian matrix is ​​assumed to be a diagonal matrix, H = diag([H 1,1 ,...,H n,n]), each element on the diagonal is a positive number; the quadratic approximation of the objective function is decomposed into:

[0077]

[0078] Among them, each w i The calculation method is:

[0079]

[0080] Among them, H i,i is the i-th element on the diagonal of the Hessian matrix; w i is the function weight.

[0081] when When the original objective function J(w) of the convolution in the i-th dimension is The contribution of w is suppressed, and the L1 regularization term i Push to 0, w i The optimal value of is 0; when When , the L1 regularization term will not change w i Push to 0 and move on the original basis

[0082] like Figure 2 As shown, when the L1-norm pruning algorithm removes m convolution kernels from the i-th convolution layer, the m weights w i The convolution kernels with a value of 0 and the corresponding feature maps are pruned, and the convolution kernels related to the pruned feature maps in the next convolution layer are also removed. The weight matrix from layer i to layer i+1 is then recreated, and the remaining weight parameters are copied to the new model. After the convolution kernels of layer i+1 are pruned, the convolution kernel parameters retained in this layer are not only effectively reduced in terms of the number of convolution kernels, but are also related to the convolution kernels pruned in layer i. Therefore, the channel dimension of each convolution layer is also reduced. After setting the pruning rate, the above pruning process is repeated for each layer. Pruning stops when the set value is reached, and the pruned model is fine-tuned to compensate for potential temporary degradation.

[0083] The FPGM (Filter Pruning via Geometric Median) algorithm ranks the importance of convolution kernels based on the evaluation criterion of geometric median, and eliminates redundant filters to reduce the number of model parameters.

[0084] If the network has K convolution layers, and the jth layer of the neural network has N convolution kernels, all the convolution kernels in each layer are regarded as a multi-dimensional Euclidean space, and each filter is a point f in the Euclidean space. i Taking a 2×2 filter as an example, the Euclidean distance between any two filters is:

[0085]

[0086] Among them, A(i,j) is the parameter coordinate in the filter; then the geometric median f of the filter * for:

[0087]

[0088] Among them, f x is the undetermined geometric median; i is the filter number; j is the corresponding value of the Conv convolution layer.

[0089] The geometric median is the center point in Euclidean space, the point with the smallest Euclidean distance to all other points. Points closer to the center are more easily represented by points farther away. If a filter is close to the center point, it can be considered that its information overlaps with other filters and its function can be replaced by other filters. Pruning it out will have minimal impact on the overall performance of the network.

[0090] The specific implementation method of the FPGM pruning algorithm is:

[0091] First, update the network weight parameters, set the epoch value and pruning rate;

[0092] Starting from the first epoch, during the training process, redundant convolution kernels are set to zero according to the geometric median and pruning rate of each layer;

[0093] The zeroed convolution kernel will be updated again in the next epoch until the epoch completes the set value. After repeated iterations, the redundant filters will eventually become sparse.

[0094] The VGG-16 network model was pruned using the L1-norm pruning algorithm and the FPGM pruning algorithm at different pruning rates. Each model was trained using the dataset used in step 2, with a random split of 70% for training and 30% for validation. Batch training was used to divide the training and validation sets into multiple batches, with a batch size of 16, an epoch of 100, an initial learning rate (LR) of 0.001, and a cross-entropy error as the loss function. Adam optimization was used.

[0095] When the pruning rate is set between 10% and 50%, the FPGM pruning algorithm achieves higher accuracy metrics than the L1-norm pruning algorithm. When the pruning rate exceeds 70%, the model accuracy begins to decline significantly due to the pruning of some important parameters. At a pruning rate of 90%, the accuracy drops to 52.0% and 77.0%, respectively. At a pruning rate of 50%, the FPGM pruning algorithm achieves a pruning accuracy of 96.8%, a 1.2% decrease from the original network. The F1 score is the same as the original network, and all accuracy metrics are slightly different from the L1-norm pruning algorithm. However, the number of parameters is 26% of the original network, and the average frame time is 78.3% of the original network, both of which are superior to the L1-norm pruning algorithm. Taking into account the number of model parameters, recognition speed, and accuracy, the FPGM_prune0.5 model achieves the best overall performance and is used for carbon trace detection inside transformers.

[0096] It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of the present invention also fall within the scope of protection of the present invention.

Claims

1. A transformer discharge carbon mark detection method based on the VGG-16 pruning model is characterized by: The following steps are involved: Step 1: The micro robot takes the original internal image of the transformer along the planned path inside the oil tank of the oil-immersed transformer and transmits it back to the control platform; Step 2: The control platform uses image enhancement and data enhancement to process the dataset constructed by the image acquired in step 1, performs image enhancement on the original dataset, and then expands the quantity of the dataset after image enhancement to obtain the dataset used in the experiment; The specific implementation method of the image enhancement in step 2 is: enhancing the image by improving the histogram equalization algorithm. The specific implementation method of the data enhancement in step 2 is: expanding the image enhanced dataset by using four methods: Gaussian blur processing, horizontal and vertical flipping, scaling, and horizontal and vertical translation. At the same time, image fusion data enhancement is performed according to the Mosaic algorithm. Improving the histogram equalization algorithm to enhance the image includes the following steps: Step 2.1, divide the image obtained in step 1 into 8 parts horizontally and 4 parts vertically; Step 2.2, calculate the histogram of each sub-block, and determine the dynamic range and emphasis factor of each histogram; New grayscale value dynamic range r i The calculation method is: c i =S max -S min Among them, c i is the gray level range of each sub-block histogram, w i is the emphasis factor of each sub-block histogram, n i The total number of pixels in each sub-block, m i is the number of pixel types in each sub-block, W k is the emphasis factor of the kth gray level, t is the total gray value level, L is the number of divided sub-blocks, S max The maximum grayscale value in each sub-block histogram; S min The minimum grayscale value in each sub-block histogram; Step 2.3, perform independent grayscale equalization on each sub-histogram in step 2.1; Perform histogram equalization on each sub-block, and its mapping function is: Among them, S i is the grayscale value of the mapped image; n is the total number of pixels in the sub-block; n k is the number of pixels of gray level k in the sub-block; Step 3. Establish a VGG-16 network model and conduct pruning comparison experiments of the L1-norm pruning algorithm and the FPGM pruning algorithm under different pruning rate settings. Based on the experimental results, the pruning algorithm model suitable for carbon mark detection inside the transformer is finally determined, and carbon mark detection is performed inside the transformer.

2. The transformer discharge carbon mark detection method based on the VGG-16 pruning model according to claim 1 is characterized in that: The specific implementation method of data enhancement for image fusion based on the Mosaic algorithm is as follows: randomly select one picture in the group with carbon mark labels for size compression, randomly select another three pictures in the entire data set for random cropping, and splice the four pictures into one picture with the classification label of "with carbon marks". In this way, several pictures with carbon mark labels are generated to enrich the picture background and reduce the scale of large carbon marks.

3. The transformer discharge carbon mark detection method based on the VGG-16 pruning model according to claim 1 is characterized in that: The specific implementation method of step 3 is: establishing a VGG-16 network model, performing network pruning with different pruning rates based on the L1-norm pruning algorithm and the FPGM pruning algorithm for the experimental data set obtained in step 2, obtaining the optimal pruning algorithm through comparison, and performing carbon mark detection on the inside of the transformer using the optimal pruning algorithm.

4. The transformer discharge carbon mark detection method based on the VGG-16 pruning model according to claim 1 is characterized in that: The specific calculation method of the L1-norm pruning algorithm in step 3 is: The objective function of the L1 norm regularization term for: Among them, J(w) is the original objective function; w is the weight matrix transferred between each layer, x i is the input vector; i is the expected output; α is a hyperparameter that weighs the relative contribution of the regularization term to the objective function J, and its range is α∈[0,∞); m is the mth parameter, Objective gradient function for: Where w* is the weight matrix caused by the regularization term; H is the Hessian matrix of J(w) at w*; the Hessian matrix is ​​a diagonal matrix, H = diag([H 1,1 ,...,H n,n ]), each element on the diagonal is a positive number; the quadratic approximation of the objective function is decomposed into: Among them, H i,i is the i-th element on the diagonal of the Hessian matrix; w i is the function weight; when When the original objective function J(w) of the convolution in the i-th dimension is The contribution of w is suppressed, and the L1 regularization term i Push to 0, w i The optimal value of is 0; when When , the L1 regularization term will not change w i Push to 0 and move on the original basis After the L1-norm pruning algorithm sets the pruning rate, it stops pruning for each layer when the set value is reached, and fine-tunes the pruned model to compensate for potential temporary degradation.

5. The transformer discharge carbon mark detection method based on the VGG-16 pruning model according to claim 1 is characterized in that: The specific method of the FPGM pruning algorithm in step 3 is as follows: updating the network weight parameters, setting the epoch value and pruning rate; starting from the first epoch, during the training process, the redundant convolution kernels are set to zero according to the geometric median and pruning rate of each layer; the zeroed convolution kernels are re-updated with parameters in the next epoch until the epoch execution completes the set value, and iterates repeatedly until the redundant filters tend to be sparse.

Images