Method, system and equipment for identifying and positioning parts of overhead line system and medium

Abstract

The invention relates to the field of engineering image intelligent recognition, in particular to a contact network part recognition and positioning method, system and device and a medium, and the method uses Faster R-CNN as a basis, combines a multi-scale balance attention feature pyramid network to carry out multi-scale feature fusion, and uses SE module as a basic attention framework to carry out multi-scale feature fusion on the basis of the Faster R-CNN. The features of different hierarchies are reconstructed into a uniform size in combination with residual connection, so that feature enhancement is realized, attention to effective features is enhanced, irrelevant features are inhibited, and the recognition precision of the model is improved; meanwhile, the penalty term is introduced on the basis of a frame coordinate regression loss function, an anchor is promoted to be closer to a real target (GT), the positioning accuracy is improved, efficient and accurate recognition and positioning of contact network parts are achieved by comprehensively applying multi-scale feature fusion, an attention mechanism and the optimized loss function, and the positioning accuracy is improved. And powerful support is provided for safety monitoring and maintenance of the overhead line system.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recognition of engineering images, and in particular to a method, system, equipment and medium for identifying and positioning contact network parts. Background Art

[0002] The overhead contact system is an important part of the railway electrification system. It is responsible for transmitting electrical energy to high-speed locomotives. The continuous operation of the overhead contact system ensures the high-speed operation of high-speed trains. Since the overhead contact system interacts with locomotives in an open-air environment for a long time, components are prone to failure. Once a component is abnormal, it poses a great threat to the safety of train operation. However, since the overhead contact components account for a small proportion of the entire image of the 2C and 4C systems, it is relatively difficult to directly detect faults in each component.

[0003] CN111753653A discloses a method for identifying and locating fasteners of high-speed railway contact network based on attention mechanism. The attention mechanism is introduced into the ResNet50 residual convolutional network, and the attention model is added to each residual module block to form a residual attention network as a convolutional neural network for feature extraction; through the regional recommendation module, the accuracy of fastener positioning is enhanced, and the accuracy rate of contact network fastener identification is more than 90%. Among them, after the Faster R-CNN model integrates the feature pyramid network, it begins to focus on the transmission of shallow detail information to deep feature maps, but does not achieve deep fusion of features and information at each layer, which needs to be improved. Summary of the invention

[0004] In view of the problem in the prior art that the recognition accuracy is poor due to the failure to achieve fusion of features and information at each layer, the present invention provides a contact network component recognition and positioning method, system, device and medium.

[0005] The present invention is achieved through the following technical solutions: A method for identifying and locating contact network components comprises the following steps: Step 1: Input the image to be tested into a multi-scale balanced attention feature pyramid network based on Faster R-CNN for feature fusion to generate four multi-scale feature maps; Step 2: Input the four multi-scale feature maps into the RPN network to select the target box; Step 3: Unify the multi-scale feature map and the target box in the ROI pooling layer to obtain a pooled feature map of fixed size; Step 4: Input the pooled feature map into the fully connected layer to obtain the feature vector; Step 5: Input the feature vector into the classification branch and the regression branch to obtain the location of the target.

[0006] Preferably, in step 1, the process of feature fusion is: Assume that the number of feature pyramid layers is , using Conv2, Conv3, Conv4 and Conv5 of the feature extraction network as output; according to the definition of FPN, firstly, the outputs of four different scales are generated into feature maps with the same size as the corresponding layer, denoted as ,right Perform multi-scale feature fusion through upsampling and downsampling to generate a fused feature map .

[0007] Preferably, the upsampling operation selects the interpolation method, and the downsampling operation uses a 3*3 convolution kernel with a step size of 2.

[0008] Preferably, in step 1, the multi-scale balanced attention feature pyramid network based on Faster R-CNN uses SEmodule as the basic attention framework, combines residual connection, and uses the attention mechanism to enhance the effective features of the fused feature map.

[0009] Preferably, the process of fusion feature map enhancement is: First, the features at different levels are Reconstructed by interpolation or maximum pooling The size of , then take the average value to get the balanced semantic feature ; The fused balanced semantic features As the input of Attention Module, generate the corresponding attention map; In order to feed back the balanced semantic feature information to each level, the output of the Attention Module Using the same but reverse process, reconstruct to the same size as each level, we get ,and Perform the addition operation and get ;.

[0010] Preferably, in step 2, the target box selection process is as follows: four scale feature maps are input into the RPN network, the RPN network first uses a sliding frame to traverse the feature map, generates an initial anchor frame and sends it to the classification branch and regression branch of the RPN network respectively to realize the judgment of whether the target is a background class and the preliminary correction of the bounding box; finally, through non-maximum suppression (NMS), a corresponding number of target boxes are selected and their position information is recorded.

[0011] Preferably, in the target box selection and regression branch of the fully connected layer, based on the bounding box coordinate regression loss function, a center point distance penalty term is introduced to make the anchor closer to the GT.

[0012] A contact network component identification and positioning system comprises a data acquisition module, a data processing module and a data output module. The data acquisition module is used to acquire an image to be tested, the data processing module is used to acquire multi-scale features, select a target frame, perform pooling layer unification and acquire a feature vector, and the data output module is used to output the monitoring position of the target.

[0013] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the contact network component identification and positioning method when executing the computer program.

[0014] A storage medium stores a computer program, which, when executed by a processor, implements the steps of the contact network component identification and positioning method.

[0015] Compared with the prior art, the present invention has the following beneficial effects: A contact network component identification and positioning method of the present invention uses a multi-scale balanced attention feature pyramid network to fuse the features of each level on the basis of an FPN backbone network, and introduces an attention mechanism to enhance the semantic feature graphs integrated at each level, thereby obtaining the main features and reducing the redundant information of the feature graphs. In turn, it acts on the features of the corresponding levels to balance the differences between the levels, while enhancing the expression of useful features.

[0016] Furthermore, the Attention Module uses the attention mechanism to strengthen the effective features of the fused feature map. By extracting the attention of the channel dimension of the feature map and focusing on the most significant part of the information, the feature map involved in the prediction is fully utilized to fully utilize the original image information, which can effectively enhance the feature representation of balanced semantic features.

[0017] Furthermore, the center point rectangular box loss (CR Loss) based on the SmoothL1 loss function is introduced in the target box selection and the regression branch of the fully connected layer. By optimizing the area of ​​the rectangle composed of the anchor and the GT center points as vertices, the distance between the two center points is directly minimized, so that the anchor can quickly move closer to the GT. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of an improved Faster R-CNN combining MBA-Net and CR Loss in a method for identifying and locating contact network parts of the present invention; Figure 2It is the MBA-Net network structure in the present invention; Figure 3 Schematic diagram of the multi-scale feature fusion process in the present invention; Figure 4 It is a schematic diagram of the Attention Module processing process in the present invention; Figure 5 It is a schematic diagram of the CR Loss principle in the present invention; Figure 6 It is the PASCAL VOC dataset format in the embodiment; Figure 7 It is the LabelImg labeled 2C image in the embodiment; Figure 8 It is the LabelImg labeled 4C image in the embodiment; Fig. 9 This is the effect diagram of the identification and positioning of the hanging string on the 2C system image; Fig.10 This is an effect diagram for identifying the main components in the 4C system image. DETAILED DESCRIPTION

[0019] The present invention is further described in detail below in conjunction with specific embodiments, which are intended to explain the present invention rather than to limit it.

[0020] The present invention discloses a method for identifying and locating contact network components. Figure 1 , including the following steps: Step 1: Input the image to be tested into the multi-scale balanced attention feature pyramid network based on Faster R-CNN for feature fusion to generate four multi-scale feature maps. Specifically, input the image into the ResNet basic network to extract image features, obtain the outputs of four different convolution modules Conv2, Conv3, Conv4, and Conv5 as the feature pyramid for multi-scale prediction, and implement feature fusion processing of the feature pyramid through MBA-Net to generate four scale feature maps.

[0021] Multi-scale balanced attention feature pyramid network (MBA-Net), the structure is as follows Figure 2 As shown in the figure. Based on the FPN backbone network, MBA-Net fully integrates the features of each layer and introduces the attention mechanism to act on the semantic feature graphs integrated at each layer. By extracting the main features and reducing the redundant information of the feature graph, and in turn acting on the features of the corresponding layers, the differences between the layers are balanced and the expression of useful features is enhanced.

[0022] MBA-Net is based on FPN and fully integrates the multi-scale feature maps. MBA-Net uses ResNet as the basic network for feature extraction and uses the weights trained by ImageNet as the pre-training model. Assume that the number of feature pyramid layers is , using Conv2, Conv3, Conv4 and Conv5 of the feature extraction network as output; according to the definition of FPN, the outputs of four different scales are used to generate feature maps with the same size as the corresponding layer, denoted as ,right Perform multi-scale feature fusion (as shown in the solid box in Figure 2) to generate a fused feature map . Fusion feature map Each scale feature map in All feature maps in are related, for details, refer to Figure 3 In this invention, upsampling and downsampling are used to Refactored to The same size is used for fusion operation. The upsampling operation uses interpolation method, and the downsampling operation uses 3*3 convolution kernel with a step size of 2. Fusion feature map The imbalance of position information and semantic information in feature maps of each scale is balanced.

[0023] In order to reduce the information redundancy of feature maps at each scale and further enhance feature expression, this paper adopts SE module as the basic attention framework and adds residual connection, which is called Attention Module. The structure is as follows: Figure 5 As shown in Figure 2, the Attention Module uses the attention mechanism to enhance the effective features of the fused feature map. Figure 2 As shown in the dotted box in , by extracting the attention of the channel dimension of the feature map and focusing on the most significant part of the information, the feature representation of balanced semantic features can be effectively enhanced.

[0024] Specifically, we first combine the features of different levels Reconstructed by interpolation or maximum pooling The size of , then take the average value to get the balanced semantic feature . The index of the layer feature is denoted by , the index of the lowest and highest level features is denoted by , The formula is as follows:

[0025] The fused balanced semantic features As the input of the Attention Module, the corresponding attention map is generated. The spatial information of the feature map is aggregated by global average pooling to generate a spatial context descriptor: , a multi-layer perceptron (MLP) is used to generate a channel attention map The size of the hidden layer in MLP is set to , is the reduction ratio. The formula for the above process is as follows:

[0026]

[0027] here represents the sigmoid function, represents the weight parameter of the MLP in the channel attention branch. After the above operation, this paper will get the attention map Input feature map for Attention Module , and obtain the feature representation enhanced by the attention mechanism , the formula is as follows:

[0028] In order to feed back the balanced semantic feature information to each level, the output of the Attention Module Using the same but reverse process, reconstruct to the same size as each level, we get ,and Perform the addition operation and get The process formula is as follows:

[0029] Comparison Compared with FPN, it balances the differences between layers, enhances the original features of each layer, and reduces the information redundancy of feature maps at each scale. The subsequent object detection process is consistent with FPN.

[0030] Step 2: Input the four multi-scale feature maps into the RPN network to select the target frame. The target frame selection process is as follows: input the four scale feature maps into the RPN network. The RPN network first uses a sliding frame to traverse the feature map, generates an initial anchor frame and sends it to the classification branch and regression branch of the RPN network to determine whether the target is a background class and make a preliminary correction to the bounding box. Finally, through non-maximum suppression (NMS), a corresponding number of target frames are selected and their location information is recorded.

[0031] When selecting the target box, based on the bounding box coordinate regression loss function, a center point distance penalty term is introduced to make the anchor close to the GT. Specifically, based on the SmoothL1 loss function, a center point rectangular box loss (CR Loss) is designed. By optimizing the area of ​​the rectangle composed of the anchor and the GT center point as vertices, the distance between the two center points is directly minimized, so that the anchor quickly moves closer to the GT, such as Figure 5 As shown in the figure. CR Loss adds a center point distance penalty term based on the bounding box coordinate regression loss function. The loss function formula is as follows:

[0032] here , Represent the center points of the prediction box and the annotation box respectively, and Represented by , The area of ​​the rectangle formed by two points, Indicates the smallest rectangular box that can accommodate both the prediction box and the annotation box.

[0033] Step 3: Unify the multi-scale feature map and the target box in the ROI pooling layer to obtain a pooled feature map of fixed size; Step 4: Input the pooled feature map into the fully connected layer to obtain the feature vector; Step 5: Input the feature vector into the classification branch and the regression branch to obtain the position of the target. The classification branch determines the specific target category of the object in the target frame, and the regression branch accurately corrects the position coordinate information of the target frame. The output information of the comprehensive classification branch and the regression branch is used to achieve accurate detection of the target in the image.

[0034] A contact network component identification and positioning method of the present invention uses Faster R-CNN as a basis, combines a multi-scale balanced attention feature pyramid network to perform multi-scale feature fusion, uses SE module as a basic attention framework, and combines residual connection to reconstruct features of different levels into a uniform size, thereby achieving feature enhancement, strengthening the focus on effective features, suppressing irrelevant features, and improving the recognition accuracy of the model; at the same time, the penalty term is introduced on the basis of the bounding box coordinate regression loss function, so as to prompt the anchor to be closer to the real target (GT), thereby improving the positioning accuracy, and realizing efficient and accurate identification and positioning of contact network components through the comprehensive use of multi-scale feature fusion, attention mechanism and optimized loss function, thereby providing strong support for the safety monitoring and maintenance of the contact network.

[0035] Example PASCAL VOC2012 and MS COCO2014 are used as validation datasets for the performance of the method. PASCAL VOC2012 has 21 classes, and uses a train set for training, which contains 5717 images; a val set for testing, which contains 5823 images. MS COCO2014 has 81 classes, and uses a train set for training, which contains 82783 images; a val set for testing, which contains 40504 images.

[0036] 1465 and 1600 high-resolution contact network images were selected from the 2C system and 4C system of the high-speed railway contact network for engineering experiments. They were made into PASCAL VOC format datasets, where 2C images are used for the identification and positioning of hanging strings, and 4C images are used for the identification and positioning of supporting device parts. The ratio of training set to test set is 4:1.

[0037] The PASCAL VOC dataset format is as follows Figure 6 As shown in FIG. 1 , this embodiment only focuses on the Annotations, ImageSets, and JPEGImages folders required for target detection. Among them, the JPEGImages folder stores the source images, namely 1465 2C contact network images (such as Figure 7 ) and 1600 4C overhead line images (such as Figure 8 ); XML files are stored in the Annotations file, corresponding to the pictures in JPEGImages. The XML annotation files mainly contain information such as the category annotation and location coordinates of the target; the Main subfolder of the ImageSets folder stores txt files. These txt files divide the pictures into three sets: train, val and test according to the file name. For example, the train.txt under Main records the set used for training.

[0038] The LabelImg visual image annotation tool is used to annotate the 2C system images and 4C system images of the high-speed railway overhead contact network. For the 2C image, only the suspension string is annotated with the label 0. For the 4C image, seven types of components, including the arm base, insulator, U-bolt, locator support, double-sleeve connector, diagonal support sleeve and locating ring connector, are annotated with labels 1-7 respectively.

[0039] After annotating each image in JPEGImages, an XML annotation file with the same name as the annotated image will appear in the Annotations folder to record the category, location information, etc. of the annotated target.

[0040] Computing environment: CPU is Xeon E5-2664 v4 Gold, main frequency is 3.2GHz, 16 cores; GPU is NVIDIA TeslaP100, 16GB video memory, 3584 cores. Deep learning open source framework Pytorch1.5.0 is used for deep network model training and testing.

[0041] In the performance experiment of the PASCAL VOC2012 dataset, this embodiment uses the improved Faster R-CNN as the basic detector and ResNet as the feature extraction network to evaluate the proposed model. The model is trained using 5717 images and tested using 5823 images.

[0042] First, to verify the effectiveness of MBA-Net, a comparative test was conducted between MBA-Net and FPN, and the IoU threshold was set to 0.5. Table 1 shows that the MAP value has been improved accordingly on the three depths of ResNet, ResNet50, ResNet101 and ResNet152 by 1.3%, 0.9% and 0.7% respectively.

[0043] In order to understand the details of MAP improvement in detail, this embodiment selects 5 types of small targets from 20 categories, and the AP results of each category are shown in Table 2. In the 5 small target categories, MBA-Net has different degrees of improvement. Compared with FPN, the experimental results of MBA-Net show good performance improvement, which illustrates the effectiveness of the attention mechanism used in the present invention for the fusion of FPN layers.

[0044] Table 1 Experimental results on PASCAL VOC2012 dataset

[0045] Table 2 Experimental results of small objects in PASCAL VOC2012 dataset

[0046] To test the performance of CR Loss, MBA-Net and CR Loss were combined and tested on the PASCAL VOC2012 dataset in this embodiment, and the results are shown in Table 3. First, without MBA-Net, the performance of SmoothL1Loss and CR Loss were compared. ResNet50 and ResNet101 using CR Loss were improved by 0.3% and 0.4% respectively. After combining with MBA-Net, the detector performance was further improved. ResNet50 + MBA-Net + CR Loss improved by 1.5% relative to ResNet50, reaching 72.9%, and ResNet101 + MBA-Net + CR Loss improved by 1.2% relative to ResNet101, reaching 74.4%.

[0047] Table 3 Experimental results of MBA-NET and CR Loss combined on PASCAL VOC2012 dataset

[0048] To further verify the performance of the proposed method, the model was tested on the MS COCO2014 dataset in this embodiment. The MS COCO2014 dataset contains 80 classes and more than 80,000 images for training, which can be a good way to test the performance of the detector. This embodiment uses the train set for training and the val set for testing. The experiment uses the improved Faster R-CNN and tests it on the ResNet network. The purpose of this experiment is to test the impact of the combination of MBA-Net and CR loss on the entire detection network, so any performance improvement can prove its contribution to better performance. Table 4 describes the performance of the detector using ResNet50 and ResNet101 on the validation set. ResNet50 combined with MBA-Net and CR Loss is 1.5% higher than ResNet50 alone on MAP@IOU=0.5, 1.4% higher on MAP@IOU=0.7, and 1.1% higher on MAP@IOU=0.5:0.95. MBA-Net and CR Loss also improve the model on the ResNet101 network, and compared with the basic ResNet101, the improvement is 1.3% on MAP@IOU=0.5 and 1.3% on MAP@IOU=0.7. In general, this experiment describes the role of the proposed method in the Faster RCNN target detection network, showing a significant performance improvement.

[0049] Table 4 Experimental results on MS COCO2014 dataset

[0050] The hanging string recognition and positioning experiment was conducted on the 2C system dataset. In the experiment, this section used three different detectors for comparative tests, including the improved Faster R-CNN, RetinaNet and SSD. Since the dataset has high pixels and the detection targets are small, this section uses SSD512 instead of the faster SSD300. The performance of the test model is shown in Table 5. Overall, in the entire test, the improved Faster R-CNN showed obvious advantages in test accuracy, among which the performance on ResNet101+ MBA-Net + CR Loss was the best, with 86.8% and 83.9% on MAP@IOU=0.5 and MAP@IOU=0.7 respectively. ResNet50 combined with MBA-Net and CR Loss also has a certain improvement over ResNet50. RetinaNet performed best on ResNet101, with 78.8% and 72.7% on MAP@IOU=0.5 and MAP@IOU=0.7 respectively. Compared with the improved Faster R-CNN and RetinaNet, SSD has an input size of 512*512 and is faster, but performs poorly in terms of accuracy.

[0051] Table 5 Experimental results of the hanging string dataset

[0052] To further illustrate the good performance of the proposed method in the task of cable detection, this paper uses the models trained by different detectors on the contact network dataset to perform cable identification and positioning on the 2C system image, and displays the effect diagram in Fig. 9 The visualization results show that the improved Faster R-CNN shows the best detection effect, which is significantly better than RetinaNet. The visualization results also reflect the feasibility of the proposed method in the detection task of hanging string engineering.

[0053] Identify and locate 4C data ResNet101 is used as the basic feature extraction network, and the detection categories are arm base, insulator, U-bolt, locator support, double sleeve connector, diagonal support sleeve and positioning ring connector, corresponding to seven categories 1-7 respectively. The experimental results are shown in Table 6. The recall values ​​of the seven categories basically reached 100%, achieving nearly full coverage of the annotation box, which can ensure that missed detection is avoided as much as possible in actual detection tasks. Good detection performance overcomes the shortcomings of missed detection and false detection during manual detection, and provides a solid guarantee for the second step of the fault detection process.

[0054] Table 3-6 4C system component identification test results

[0055] To further demonstrate the detection effect of the model, this section visualizes the detection box information in the 4C system image and shows the recognition effect of 7 types of components. The recognition effect of the main components is as follows: Fig.10 .

[0056] The present invention also discloses a contact network component identification and positioning system, including a data acquisition module, a data processing module and a data output module. The data acquisition module is used to acquire the image to be tested, the data processing module is used to acquire multi-scale features, select target frames, perform pooling layer unification and acquire feature vectors, and the data output module is used to output the monitoring position of the target.

[0057] The present invention also discloses an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the contact network component identification and positioning method when executing the computer program.

[0058] The present invention also discloses a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for identifying and locating contact network components are implemented, providing powerful technical support for the safe monitoring and maintenance of the contact network, which not only improves the accuracy and efficiency of identification and positioning, but also enhances the flexibility and scalability of the system, and has broad application prospects.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to impose any limitation on the technical solution of the present invention. Those skilled in the art should understand that, without departing from the spirit and principles of the present invention, the technical solution can also be subjected to several simple modifications and substitutions, and these modifications and substitutions are also within the scope of protection covered by the claims.

Claims

1. A method for identifying and locating contact network components, characterized in that: The following steps are involved: Step 1: Input the image to be tested into a multi-scale balanced attention feature pyramid network based on Faster R-CNN for feature fusion to generate four multi-scale feature maps; Step 2: Input the four multi-scale feature maps into the RPN network to select the target box; Step 3: Unify the multi-scale feature map and the target box in the ROI pooling layer to obtain a pooled feature map of fixed size; Step 4: Input the pooled feature map into the fully connected layer to obtain the feature vector; Step 5: Input the feature vector into the classification branch and the regression branch to obtain the location of the target.

2. The contact network component identification and positioning method according to claim 2 is characterized in that: In step 1, the process of feature fusion is: Assume that the number of feature pyramid layers is , using Conv2, Conv3, Conv4 and Conv5 of the feature extraction network as output; according to the definition of FPN, firstly, the outputs of four different scales are generated into feature maps with the same size as the corresponding layer, denoted as ,right Perform multi-scale feature fusion through upsampling and downsampling to generate a fused feature map .

3. The contact network component identification and positioning method according to claim 2, characterized in that: During reconstruction, the interpolation method is selected for upsampling; during reconstruction, a 3*3 convolution kernel with a step size of 2 is used for downsampling.

4. The contact network component identification and positioning method according to claim 2, characterized in that: In step 1, the multi-scale balanced attention feature pyramid network based on FasterR-CNN uses SE module as the basic attention framework, combined with residual connection, and uses the attention mechanism to enhance the effective features of the fused feature map.

5. The contact network component identification and positioning method according to claim 4, characterized in that: The process of fusion feature map enhancement is: First, the features at different levels are Reconstructed by interpolation or maximum pooling The size of , then take the average value to get the balanced semantic feature ; The fused balanced semantic features As the input of Attention Module, generate the corresponding attention map; In order to feed back the balanced semantic feature information to each level, the output of the Attention Module Using the same but reverse process, reconstruct to the same size as each level, we get ,and Perform the addition operation and get ;.

6. The contact network component identification and positioning method according to claim 2, characterized in that: In step 2, the target box selection process is as follows: the four scale feature maps are input into the RPN network. The RPN network first uses a sliding frame to traverse the feature map, generates an initial anchor frame and sends it to the classification branch and regression branch of the RPN network respectively to realize the judgment of whether the target is a background class and the preliminary correction of the bounding box; finally, through non-maximum suppression (NMS), the corresponding number of target boxes are selected and their position information is recorded.

7. The contact network component identification and positioning method according to claim 6, characterized in that: In the target box selection and regression branch of the fully connected layer, based on the bounding box coordinate regression loss function, a center point distance penalty term is introduced to make the anchor closer to the GT.

8. A contact network component identification and positioning system, characterized in that: It includes a data acquisition module, a data processing module and a data output module. The data acquisition module is used to acquire the image to be tested. The data processing module is used to acquire multi-scale features, select target frames, perform pooling layer unification and acquire feature vectors. The data output module is used to output the monitoring position of the target.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the contact network component identification and positioning method described in any one of claims 1 to 7 are implemented.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the contact network component identification and positioning method described in any one of claims 1 to 7 are implemented.

Images