Transformer component detection method based on Fast RCNN

Abstract

The invention discloses a Fast RCNN-based transformer component detection method, which comprises the following steps: replacing an original VGG16 with a feature extraction network ResNet50 to serve as a trunk feature extraction network, adopting a transfer learning mode to take a weight trained on an ImageNET data set as an initial weight of a model, and replacing a region-of-interest pooling operation with a bilinear interpolation method, so as to improve the detection accuracy of a transformer component. And carrying out clustering analysis on the original data set through a K-means clustering algorithm, adopting a contour coefficient method as an evaluation standard, and customizing an anchor frame of a transformer component fault according to a clustering result. A thought is provided for application of deep learning in power equipment fault detection, the model can be carried with various mobile detection platforms subsequently, has high engineering use value, solves the problems of insufficient feature abstraction degree, low accuracy and the like of transformer component detection images, is beneficial to saving manpower, guarantees the safety of inspection personnel, and improves the detection efficiency. And the problems of inspection blind spots and poor real-time performance caused by large subjectivity are avoided, so that the transformer component detection method with higher reliability is realized.

Description

Technical Field

[0001] The present invention relates to the field of transformer component detection methods, and in particular to a transformer component detection method based on Fast RCNN. Background Art

[0002] With the continuous development of the economy, people's living standards are improving, various household appliances and office appliances are used more frequently, and the safety and stability of the power system are facing huge challenges. The transformer is the most important electrical equipment. If a fault occurs, it will seriously affect the normal operation of the power system. In particular, various components of the transformer often reflect key information related to the operating status of the transformer. It is not only the difficulty of transformer inspection, but also the key to realizing automatic identification of transformer defects. Therefore, how to accurately detect and identify faults of various components of substation transformers is of great research significance.

[0003] At present, the main status and problems of transformer component detection methods at home and abroad are summarized as follows:

[0004] 1) Based on the image-based automatic recognition method of abnormal defects in transformer appearance, the substation transformer crack images are collected and denoised, and then the features of the substation transformer crack images are extracted. Finally, a neural network is used to establish a substation transformer crack image recognition model, which can obtain better substation transformer crack image recognition results. However, for small transformer components, this method contains very limited pixel information in the image, and it is difficult to achieve accurate recognition.

[0005] 2) Based on the method of transformer crack image recognition model in substation, the transformer crack image is classified and the categories of transformer crack images are correctly divided, so as to determine the fault category of the transformer component. However, in practical applications, the acquisition of transformer crack images is affected by weather, environment and acquisition equipment, and the image will inevitably contain certain noise, which will have an adverse effect on the accurate recognition of transformer crack images.

[0006] 3) The method of substation equipment image recognition based on joint features uses the SIFT method to extract local gradient features of the image, and realizes automatic recognition of transformer faults with the help of template images. On the basis of SIFT features, the sparse coding features of the image are added, and the substation transformer and other equipment are recognized through joint features. However, the artificially designed image features are not abstract enough, the generalization ability is generally weak, the recognition accuracy is unstable, and it is difficult to adapt to complex substation scenes.

[0007] 4) Based on the deep learning target detection model SSD method, the transformer fault recognition and location are realized on the basis of extracting the deep features of the substation image. Compared with the manually designed image features, the deep learning neural network is used to automatically extract the image features in depth, which improves the generalization ability of the model. However, the above method neither solves the problem that small parts contain too little pixel information, nor considers the correlation between the positions of various parts on the same device. When applied to transformer component recognition, its accuracy will be greatly affected.

[0008] At present, the problems with transformer component detection mainly focus on the limited pixel information contained in the image, insufficient abstraction of image features, and low accuracy.

[0009] Therefore, it is necessary to propose a transformer component detection method based on Fast RCNN to solve the above technical problems. By improving the original Fast RCNN feature extraction backbone network, the region of interest pooling operation and other links, the problem of insufficient feature abstraction and low accuracy of transformer component detection images is solved, which is conducive to saving manpower, ensuring the safety of inspection personnel, and avoiding the problems of inspection blind spots and poor real-time performance caused by high subjectivity, thereby realizing a more reliable transformer component detection method. Summary of the invention

[0010] The purpose of this section is to summarize some aspects of embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0011] Therefore, the technical problem to be solved by the present invention is.

[0012] To solve the above technical problems, the present invention provides the following technical solutions: a transformer component detection method based on Fast RCNN, characterized in that it includes: using a feature extraction network ResNet50 with fewer parameters and deeper layers to replace the original VGG16 as the backbone feature extraction network, using transfer learning to use the weights trained on the ImageNET data set as the initial weights of the model, using bilinear interpolation to replace the pooling operation of the region of interest, performing clustering analysis on the original data set through the K-means clustering algorithm, using the silhouette coefficient method as an evaluation criterion, and customizing the anchor frame of the transformer component failure based on the clustering results; compared with other target detection networks, it can better meet the reliability requirements of the power site, provide an effective and practical method for the identification of multi-scale targets and the detection of transformer component failures, and provide ideas for the application of deep learning in power equipment fault detection. The model can be subsequently equipped with various mobile detection platforms and has high engineering use value. In order to improve the generalization ability of the model, the collected training sets are enhanced by graying, flipping, translating, sharpening and other methods. In the backbone feature extraction network, the VGG16 network is replaced by the deep neural residual network ResNet50. Through the sub-module of RseNet, a shortcut branch is introduced to convert the mapping to be fitted into the residual F(x) = H(x)-x. In this way, ResNet50 can avoid information loss, gradient explosion or disappearance when extracting deeper feature information.

[0013] As a preferred method of the transformer component detection method based on Fast RCNN described in the present invention, the feature extraction network ResNet50 replaces the original VGG16 as the backbone feature extraction network, which means: according to the requirements of transformer component detection, the feature extraction network of the model is improved, and ResNet50 is used to replace VGG16, with fewer parameters, a deeper network structure, and richer feature extraction of the target, avoiding information loss, gradient explosion or disappearance when extracting deeper feature information. The full convolution network structure RPN is used to generate a corresponding anchor frame for each point on the feature map. For targets of different sizes, the feature map is framed using the anchor frame, and the RPN is then screened to further adjust a better position, and the classification prediction value and prediction offset value generated by the RPN convolutional network are applied to all anchor frames, and the suggestion frames that exceed the image size are filtered to produce the final suggestion frame.

[0014] As a preferred mode of the transformer component detection method based on Fast RCNN described in the present invention, the use of bilinear interpolation to replace the pooling operation of the region of interest means that the Fast RCNN algorithm is based on an end-to-end transformer component target detection framework, which significantly improves the accuracy and speed of transformer component target detection, adopts the idea of ​​bilinear interpolation in Mask RCNN, avoids the problem that the rounding operation will lose some information, and the RoI Align layer algorithm preserves the characteristics of the original area to the greatest extent possible, which is conducive to improving the accuracy of the model. Improve the RoIPooling layer in Fast RCNN. Since the RoIPooling layer in the original detection network is in feature calculation, the rounding operation will discard the value after the decimal point in the pixel, and the feature information loss brought about by this will affect the model's subsequent processing of information, resulting in a decrease in the overall accuracy of Fast RCNN. Therefore, the idea of ​​bilinear interpolation in Mask RCNN is adopted to replace the RoIPooling layer with the RoI Align layer, avoiding the problem that the rounding operation will lose some information. The RoI Align layer algorithm preserves the characteristics of the original area to the greatest extent possible, and improves the accuracy of the model.

[0015] As a preferred mode of the transformer component detection method based on Fast RCNN described in the present invention, the cluster analysis of the original data set by the K-means clustering algorithm and the customized anchor frame of the transformer component fault by the clustering result refers to: training a targeted anchor frame by the K-means clustering algorithm for the length and width of the target box to be detected in the original transformer component data set, re-customizing the size, aspect ratio and number of the anchor frame, so that the transformer component detection network can more easily learn accurate fault detection anchor frames and improve the detection accuracy. The K-means clustering algorithm is used to analyze the data set, and a new anchor frame is designed for the fault information of the transformer component. After statistically calculating the length and width of the target box to be detected in the original data set, a targeted anchor frame is trained by the K-means clustering algorithm, and the anchor frame is re-customized to make the model converge faster and the detection ability stronger.

[0016] As a preferred embodiment of the transformer component detection method based on Fast RCNN described in the present invention, the feature extraction network ResNet50 with fewer parameters and deeper layers is used. The result of the comparison between the residual block structure of the ordinary network and ResNet50 is that for the final target mapping of the neural network is H(x), the ordinary network needs to directly fit H(x), and the submodule proposed by ResNet50 introduces a shortcut branch to convert the mapping to be fitted into the residual F(x)=H(x)-x, so that ResNet50 can avoid information loss, gradient explosion or disappearance when extracting deeper feature information. RPN combines the K-means clustering algorithm to obtain a customized anchor frame to trim targets that exceed the image size, and obtain the final RoI. The RoI region is input into RoI Align for regional calibration pooling, and an accurate feature map is obtained and sent to RCNN. The feature map is used by RCNN to classify the faults of transformer components.

[0017] As a preferred embodiment of the transformer component detection method based on Fast RCNN described in the present invention, the idea of ​​using bilinear interpolation in Mask RCNN refers to: using bilinear interpolation to obtain values ​​whose coordinates are floating point numbers, still mapping the ROI to the feature map, but retaining floating point numbers for the coordinates and sizes, thereby improving the detection performance of small transformer components.

[0018] As a preferred embodiment of the transformer component detection method based on Fast RCNN described in the present invention, the re-customizing the anchor frame using the K-means clustering algorithm refers to: training a targeted anchor frame through the K-means clustering algorithm, re-customizing the anchor frame size, aspect ratio and quantity, so that the detection network can more easily learn accurate transformer component fault detection anchor frames, thereby improving detection accuracy.

[0019] The beneficial effects of the present invention are as follows: to deal with the problems of insufficient abstraction and low accuracy of the features of transformer component detection images, compared with other target detection networks, it can better meet the reliability requirements of the power site, provide an effective and practical method for the identification of multi-scale targets and the detection of transformer component faults, and provide ideas for the application of deep learning in power equipment fault detection. The model can be subsequently equipped with various mobile detection platforms and has high engineering use value. By improving the original Fast RCNN feature extraction backbone network, the pooling operation of the region of interest, and other links, the problems of insufficient abstraction and low accuracy of the features of the transformer component detection image are solved, which is conducive to saving manpower, ensuring the safety of inspection personnel, and avoiding the problems of inspection blind spots and poor real-time performance caused by high subjectivity, thereby realizing a more reliable transformer component detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0021] Figure 1 A Fast RCNN algorithm network framework of a transformer component detection method based on Fast RCNN according to an embodiment of the present invention;

[0022] Figure 2 A comparison diagram of the residual block structure of a common network and a ResNet50 of a transformer component detection method based on Fast RCNN according to an embodiment of the present invention;

[0023] Figure 3 A main process diagram of ROIAlign of a transformer component detection method based on Fast RCNN according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0025] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention in detail, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0027] Furthermore, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0028] Embodiment 1

[0029] Reference Figure 1, the original VGG16 is replaced by the feature extraction network ResNet50 with fewer parameters and deeper layers as the backbone feature extraction network, the weights trained on the ImageNET dataset are used as the initial weights of the model by transfer learning, the bilinear interpolation method is used to replace the pooling operation of the region of interest, the original dataset is clustered by the K-means clustering algorithm, the silhouette coefficient method is used as the evaluation criterion, and the anchor frame of the transformer component fault is customized based on the clustering results. Compared with other target detection networks, it can better meet the reliability requirements of the power site, and provides an effective and practical method for the recognition of multi-scale targets and the detection of transformer component faults, and provides ideas for the application of deep learning in power equipment fault detection. This model can be subsequently equipped with various mobile detection platforms and has high engineering use value.

[0030] The improved Fast RCNN algorithm network framework of the present invention can be divided into four parts according to network functions: ResNet50 backbone feature extraction network, region proposal network (RPN), region of interest pooling (ROI Align) and fully connected network (RCNN) module. When performing transformer component detection, first input the component image, obtain the feature map through the backbone feature extraction network in the detection network, input the feature map into the RPN layer, and output the suggestion box containing the location information through the RPN layer and then input it into the region of interest pooling layer, which pools the input to a fixed dimension and then inputs it into the RCNN fully connected network to complete the target positioning and fault classification. The overall structure of Fast RCNN is as follows: Figure 1 As shown. The original data set is clustered by the K-means clustering algorithm, and the anchor frame of the transformer component fault is customized based on the clustering result. It means: for the length and width of the target box to be detected in the original transformer component data set, the K-means clustering algorithm is used to train a targeted anchor frame, and the size, aspect ratio and number of the anchor frame are re-customized, so that the transformer component detection network can more easily learn accurate fault detection anchor frames and improve the detection accuracy. Replacing the pooling operation of the region of interest with the bilinear interpolation method means: the Fast RCNN algorithm is based on an end-to-end transformer component target detection framework, which significantly improves the accuracy and speed of transformer component target detection. The idea of ​​bilinear interpolation in Mask RCNN is adopted to avoid the problem of losing some information due to rounding operations. The RoI Align layer algorithm preserves the characteristics of the original region to the greatest extent possible, which is conducive to improving the accuracy of the model. The idea of ​​bilinear interpolation in Mask RCNN means: using bilinear interpolation to obtain values ​​with floating-point coordinates, the ROI is still mapped to the feature map, but the coordinates and size are both floating-point numbers, thereby improving the detection performance of small transformer components.

[0031] Embodiment 2

[0032] Corresponding to the Fast RCNN algorithm network framework disclosed in the first embodiment of the present invention, the second embodiment of the present invention also provides a specific implementation method of the ResNet50 backbone feature extraction network, see Figure 2 , the method comprising:

[0033] In view of the low resolution of transformer component images, the deep neural residual network ResNet50 is used to replace the VGG16 network in the original Fast RCNN. The VGG16 network becomes increasingly difficult to train as the number of convolutional layers increases, accompanied by the gradient vanishing phenomenon, and the gradient correlation returned by the deeper network becomes increasingly poor. In order to solve the above problems and extract deeper feature information, the deep residual framework of the ResNet50 network is introduced. The deep residual network is deeper and has fewer parameters, so it can extract and process richer feature information. The ordinary network is compared with the ResNet50 residual block structure, as shown in the figure. Figure 2 As shown in the figure, when the final target mapping of the neural network is H(x), the ordinary network needs to directly fit H(x). The submodule proposed by ResNet50 introduces a shortcut branch to convert the mapping to be fitted into the residual F(x) = H(x)-x. In this way, ResNet50 can avoid information loss, gradient explosion or disappearance when extracting deeper feature information.

[0034] Embodiment 3

[0035] Corresponding to the Fast RCNN algorithm network framework disclosed in the first embodiment of the present invention, the third embodiment of the present invention further provides a specific implementation method of the region proposal network RPN, which includes:

[0036] RPN is a fully convolutional network structure, and its function is to generate better suggestion boxes. In the RPN structure, the sliding window will perform a convolution operation on the feature map and input the generated results into the regression layer and the classification layer. The anchor frame used in the original Fast RCNN algorithm is a priori frame designed manually in advance. It cannot adapt well to the target that needs to be detected by the transformer component of the present invention, and will eventually lead to poor detection effect. The present invention trains targeted anchor frames through the K-means clustering algorithm, and re-customizes the size, aspect ratio and number of the anchor frames, so that the detection network can more easily learn accurate fault detection anchor frames for transformer components, thereby improving detection accuracy. The K-means clustering algorithm is an unsupervised machine learning method, and the specific steps of the algorithm are as follows:

[0037] 1. Randomly select i samples from the sample set x as the initial mean vector (μ 1 , μ 2 , ..., μ i ),make

[0038] 2. Calculate all samples x i With each mean vector μ i Distance: d i =‖x i -μ i ‖;

[0039] 3. Cluster labeling based on the nearest mean vector: λ i =min(d i );

[0040] 4. Sample x i Assign to the corresponding cluster and calculate the new mean vector Update the current mean vector;

[0041] 5. Repeat steps 2-4 until all mean vectors no longer change;

[0042] 6. Output cluster partition C = {C 1 ,C 2 ,...,C i}, and get the appropriate anchor box.

[0043] The classification prediction value and prediction offset value generated by the RPN convolutional network are applied to all anchor boxes, and the suggestion boxes that exceed the image size are filtered. The overlapping boxes are removed through non-maximum suppression to obtain the final suggestion boxes. Then the suggestion boxes are filtered to generate regions of interest. When training the model, the generated suggestion boxes need to be further filtered to obtain the default number of 256 regions of interest.

[0044] Embodiment 4

[0045] Corresponding to the Fast RCNN algorithm network framework disclosed in the first embodiment of the present invention, the fourth embodiment of the present invention further provides a specific implementation method of the region of interest pooling (ROI Align) and the fully connected network (RCNN) module, which includes:

[0046] ROI Align is a regional feature aggregation method proposed in the Mask RCNN framework in instance segmentation. The bilinear interpolation method is used to replace the ROIPooling layer to retain the features of the original region to the greatest extent. The main process of using ROI Align is as follows: Figure 3 As shown, bilinear interpolation is used to obtain floating-point values ​​of coordinates. The ROI is still mapped to the feature map, but the coordinates and sizes are retained as floating-point numbers, thereby improving the detection performance of transformer widgets.

[0047] The region of interest pooling part matches the feature map and the region of interest output by the RPN and pools them to a fixed dimension to meet the processing requirements of the subsequent fully connected network. The fully connected layer and softmax are used to further fine-tune the position of the regression box of the region of interest whose features are pooled to a fixed dimension, and the detected transformer component faults are classified to obtain the final category and regression prediction. The specific steps are to divide the region of interest into regions of equal size, find the maximum pixel value in each region, and then copy and retain these values ​​to the Maxpooling layer. All regions of interest are converted to fixed dimensions and then forward propagated. The fully connected layer classifies and regresses the information after uniform pooling to a fixed dimension, and outputs the final transformer component detection results.

[0048] The original VGG16 is replaced by the feature extraction network ResNet50 with fewer parameters and deeper layers as the backbone feature extraction network. The weights trained on the ImageNET dataset are used as the initial weights of the model by transfer learning. The pooling operation of the region of interest is replaced by the bilinear interpolation method. The original dataset is clustered by the K-means clustering algorithm. The silhouette coefficient method is used as the evaluation criterion, and the anchor frame of the transformer component fault is customized based on the clustering results. In the backbone feature extraction network, a shortcut branch is introduced through the submodule of RseNet, so that ResNet50 can avoid information loss, gradient explosion or disappearance when extracting deeper feature information. The K-means clustering algorithm is used to analyze the dataset, and a new anchor frame is designed for the fault information of the transformer component. After statistically calculating the length and width of the target box to be detected in the original dataset, the K-means clustering algorithm is used to train a targeted anchor frame. The re-customization of the anchor frame makes the model converge faster and has stronger detection capabilities. The classification prediction value and prediction offset value generated by the RPN convolutional network are applied to all anchor boxes, and the suggestion boxes that exceed the image size are filtered out to produce the final suggestion boxes. The RoI Pooloing layer in Fast RCNN is improved, and the idea of ​​bilinear interpolation in Mask RCNN is adopted to replace the RoI Poolin layer with the RoI Align layer, avoiding the problem of losing some information due to rounding operations. The RoI Align layer algorithm preserves the characteristics of the original area as much as possible and improves the accuracy of the model. The RPN is combined with the K-means clustering algorithm to obtain a customized anchor box to trim the target that exceeds the image size to obtain the final RoI. The RoI area is input into RoI Align for regional calibration pooling, and the accurate feature map is obtained and sent to RCNN. The feature map is used by RCNN to detect transformer components and classify faults.

[0049] In summary, compared with other target detection networks, it can better meet the reliability requirements of power sites, provide an effective and practical method for multi-scale target recognition and transformer component fault detection, and provide ideas for the application of deep learning in power equipment fault detection. This model can be subsequently equipped with various mobile detection platforms and has high engineering use value. Solving the problems of insufficient feature abstraction and low accuracy of transformer component detection images is conducive to saving manpower, ensuring the safety of inspection personnel, and avoiding the problems of inspection blind spots and poor real-time performance caused by high subjectivity, thereby realizing a more reliable transformer component detection method.

[0050] It is important to note that the construction and arrangement of the present application shown in a number of different exemplary embodiments are illustrative only. Although only a few embodiments are described in detail in this disclosure, it should be readily understood by those who refer to this disclosure that many modifications are possible (e.g., the size, scale, structure, shape and proportion of various elements, and parameter values, etc.) without substantially departing from the novel teachings and advantages of the subject matter described in this application.

[0051] Additionally, in order to provide a concise description of exemplary embodiments, all features of an actual embodiment (ie, those features that are not relevant to the best mode presently contemplated for carrying out the invention or those that are not relevant to implementing the invention) may not be described.

[0052] It will be appreciated that in the development of any actual implementation, as in any engineering or design project, numerous implementation-specific decisions may be made. Such a development effort may be complex and time-consuming, but will be a routine task of design, fabrication, and production for those of ordinary skill having the benefit of this disclosure without undue experimentation.

[0053] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A transformer component detection method based on Fast RCNN, characterized in that: include: The feature extraction network ResNet50 replaces the original VGG16 as the backbone feature extraction network. The weights trained on the ImageNET dataset are used as the initial weights of the model by transfer learning. The bilinear interpolation method is used to replace the pooling operation of the region of interest. The original dataset is clustered by the K-means clustering algorithm. The silhouette coefficient method is used as the evaluation criterion, and the anchor frame of the transformer component fault is customized according to the clustering results.

2. The transformer component detection method based on Fast RCNN according to claim 1, characterized in that: The feature extraction network ResNet50 replaces the original VGG16 as the backbone feature extraction network, which means that according to the requirements of transformer component detection, the feature extraction network of the model is improved, and ResNet50 is used to replace VGG16, which has fewer parameters, a deeper network structure, and richer feature extraction of the target, and avoids information loss, gradient explosion or disappearance when extracting deeper feature information.

3. The transformer component detection method based on Fast RCNN according to claim 1, characterized in that: The use of bilinear interpolation to replace the pooling operation of the region of interest means that the Fast RCNN algorithm is based on an end-to-end transformer component target detection framework, which significantly improves the accuracy and speed of transformer component target detection. The idea of ​​bilinear interpolation in Mask RCNN is adopted to avoid the problem of losing some information due to rounding operations. The RoI Align layer algorithm preserves the characteristics of the original area to the greatest extent possible, which is conducive to improving the accuracy of the model.

4. The transformer component detection method based on Fast RCNN according to claim 1, characterized in that: The clustering analysis of the original data set by using the K-means clustering algorithm and the customization of the anchor frame of the transformer component fault according to the clustering result refers to: training a targeted anchor frame by using the K-means clustering algorithm for the length and width of the target box to be detected in the original transformer component data set, and re-customizing the size, aspect ratio and number of the anchor frame, so that the transformer component detection network can more easily learn accurate fault detection anchor frames, thereby improving the detection accuracy.

5. The transformer component detection method based on Fast RCNN according to claim 2, characterized in that: The result of comparing the feature extraction network ResNet50, the ordinary network and the residual block structure of ResNet50 is that for the final target mapping of the neural network is H(x), the ordinary network needs to directly fit H(x), and the submodule proposed by ResNet50 introduces a shortcut branch to convert the mapping to be fitted into the residual F(x) = H(x)-x. In this way, ResNet50 can avoid information loss, gradient explosion or disappearance when extracting deeper feature information.

6. The transformer component detection method based on Fast RCNN according to claim 3, characterized in that: The idea of ​​using bilinear interpolation in Mask RCNN is to use bilinear interpolation to obtain floating-point values ​​of coordinates, and still map the ROI to the feature map, but the coordinates and sizes are retained as floating-point numbers, thereby improving the detection performance of transformer widgets.

7. The transformer component detection method based on Fast RCNN according to claim 4, characterized in that: Re-customizing the anchor frame using the K-means clustering algorithm means: training a targeted anchor frame through the K-means clustering algorithm, re-customizing the size, aspect ratio and number of the anchor frame, so that the detection network can more easily learn the accurate fault detection anchor frame of the transformer component, thereby improving the detection accuracy.

Images