Ground penetrating radar small target B-scan identification method and system for underground facilities of transformer substation

Abstract

The invention discloses a ground penetrating radar small target B-scan identification method and system for underground facilities of a transformer substation, and relates to the field of ground penetrating radar data processing. The problems of false detection and high omission ratio caused by a large number of small target samples in underground facility ground penetrating radar detection in a power grid and the problems that a traditional image processing method depends on manual identification and labeling, and the data processing accuracy and processing speed are difficult to improve are solved. The method comprises the following steps: scanning an underground facility along a motion direction according to a UWB ultra-wideband radar system to obtain a B-scan graph; identifying the B-scan image by adopting a LableImage, and obtaining a characteristic waveform in the B-scan image; a GPR-Bfnet recognition model is constructed, and the GPR-Bfnet recognition model is trained according to the training set; and the trained GPR-Bfnet identification model is used to carry out identification and classification on the characteristic waveform in the B-scan image. The method is applied to the fields of engineering, geological exploration and military.

Description

Technical Field

[0001] The invention relates to the field of ground penetrating radar data processing, and in particular to a ground penetrating radar small target B-scan recognition method for underground facilities of a substation. Background Art

[0002] With the popularization of very large scale integrated circuit systems (VLSI System), the performance of various microprocessors including DSP has been greatly improved. Therefore, the research on ground penetrating radar, which has high requirements for computer performance, has gradually become one of the hot spots in the field of radar research. As an effective tool that uses the reflection and scattering of electromagnetic waves at the discontinuity of the medium to achieve the identification and positioning of targets less than 50m underground, the advantages of ground penetrating radar (GPR) detection are its fast speed, continuous process, high imaging clarity, and flexible and convenient operation, non-destructive testing, etc. Therefore, it has been widely used in engineering, geological exploration, military and other fields in recent years. As early as the beginning of the 20th century, the Germans began to use ground penetrating radar to detect underground metal objects. Later, as a new technology for non-destructive testing, the application of ground penetrating radar has developed from the initial detection of the thickness of the ice layer of weak attenuation media to the fields of engineering detection and geological exploration.

[0003] In the process of detecting underground facilities in substations based on ground penetrating radar, since the cross-section of underground facilities is generally less than 5cm2, the proportion of small targets in the data set is relatively large. Since small objects occupy very few pixels, they are easily disturbed by complex backgrounds and large objects, and are prone to errors or missed detections. Traditional image processing methods rely on manual recognition and annotation, which makes it difficult to improve the accuracy and speed of data processing. Summary of the invention

[0004] The present invention aims to solve the problem of high false detection and missed detection rate caused by a large number of small target samples in ground penetrating radar detection of underground facilities in power grids and the problem that the accuracy and processing speed of data processing are difficult to improve due to the reliance on manual recognition and labeling by traditional image processing methods. A ground penetrating radar small target B-scan recognition method for underground facilities in substations is proposed. The method comprises:

[0005] S1: Scan the underground facilities along the moving direction according to the UWB ultra-wideband radar system to obtain a B-scan image;

[0006] S2: Label the B-scan image using LabelImage to obtain the characteristic waveform in the B-scan image;

[0007] S3: Build a GPR-Bfnet recognition model and train the GPR-Bfnet recognition model based on the training set;

[0008] S4: Use the trained GPR-Bfnet recognition model to identify and classify the characteristic waveforms in the B-scan image.

[0009] Furthermore, a preferred embodiment is proposed, wherein step S1 further comprises:

[0010] The B-scan image is enhanced, and new images are generated from the original B-scan image through random cropping, random rotation and mosaic enhancement to construct a data set;

[0011] The data set is divided into training set, test set and validation set according to the ratio of 7:2:1.

[0012] Furthermore, a preferred method is proposed, characterized in that the GPR-Bfnet recognition model is constructed in step S3, including:

[0013] Extract low-frequency and high-frequency features through the GPR-Trans module;

[0014] Combining semantic information and location information to process features of different sizes through the GPR-Trans-BiFPN structure;

[0015] Use the bidirectional feature pyramid structure and the fast normalized fusion layer FastNormalizedFusion to fuse features of different sizes;

[0016] Use the COC-Adown module for downsampling.

[0017] Furthermore, a preferred method is proposed, wherein the GPR-Trans module includes a dual-branch information processing architecture, and a maximum pooling layer is used after the normalization layer to process low-frequency information:

[0018] X global =Attention(Q g ,Pool(K g ),Pool(V g ))

[0019] Among them, Qg represents the global query vector, Kg represents the global key vector, Vg represents the global value vector, and Xglobal represents the vector of the global branch;

[0020] Use depth-wise separable convolution DWConv to process high-frequency information;

[0021] Q,K,V=FC(X in )

[0022] V s =DWconv(V)

[0023] Q l =DWconv(Q)

[0024] K l =DWconv(K)

[0025]

[0026] X local =Attn⊙V s

[0027] Among them, Q l represents the local branch query vector, K l represents the local branch key vector, V s represents a local numerical vector, X in represents the original input vector, Q represents the query vector, K represents the key vector, V represents the numerical vector, DWconv represents the depth-wise convolution, A ttnt represents the attention mechanism, A ttn Indicates the addition of nonlinear activation function attention mechanism, X local Represents local branch output.

[0028] Furthermore, a preferred method is proposed, wherein the bidirectional feature pyramid structure and the fast normalized fusion layer Fast NormalizedFusion perform feature fusion of different sizes, including:

[0029] Establish two pathways, top-down and bottom-up; the top-down pathway is used for feature fusion, and the bottom-up pathway is used for transmitting position information;

[0030] Depthwise separable convolution is used for feature fusion, and batch normalization and activation are added after each convolution.

[0031] Furthermore, a preferred method is proposed, in which the COC-Adown channel stacking convolution downsampling replaces the traditional pooling layer, divides the image into sub-scales, and downsamples by channel stacking, and finally uses single-step convolution to optimize the amount of calculation.

[0032] Based on the same inventive concept, the present invention also proposes a ground penetrating radar small target B-scan recognition system for underground facilities of substations, the system comprising:

[0033] A scanning unit, used for scanning the underground facilities along the moving direction according to the UWB ultra-wideband radar system to obtain a B-scan map;

[0034] A characteristic waveform acquisition unit, used to label the B-scan image using LabelImage, and acquire a characteristic waveform in the B-scan image;

[0035] A model building unit, used for building a GPR-Bfnet recognition model and training the GPR-Bfnet recognition model according to a training set;

[0036] The recognition unit is used to recognize and classify the characteristic waveform in the B-scan image using the trained GPR-Bfnet recognition model.

[0037] Based on the same inventive concept, the present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the ground penetrating radar small target B-scan identification method for underground facilities of substations according to any one of the above-mentioned methods.

[0038] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the ground-penetrating radar small target B-scan identification method for underground facilities of substations as described in any of the above items.

[0039] The present invention is beneficial in that:

[0040] In the B-scan recognition method for small targets of ground penetrating radar for underground facilities of substations proposed in the present invention, an information-rich small target specialized model (GPR-Bfnet) facing the GPR data set is constructed. The model aims to enrich the information and reduce feature loss by using rich context-aware information and removing pooling in downsampling, thereby reducing the probability of missed detection and false detection of the model when facing small target samples and improving detection accuracy. Specifically, the idea of ​​Transformer is integrated into the backbone part, so that the image is output in multiple levels in the backbone and combined with Bifpn Neck for feature fusion to enrich the feature information obtained by the model. In addition, a downsampling method that reduces the image size by channel stacking is designed, and the feature information loss is reduced by fusion with the input image features. Finally, for small target detection, a bounding box regression algorithm that considers the shape and size of the bounding box itself is used to solve the problem of low detection accuracy of the GPR data set of underground facilities of substations from three aspects. Experimental results show that the proposed model improves by 6.04%, 7.87% and 1.27% in map0.5, precision and recall respectively compared with YOLOv9. In the dataset composed of small target samples extracted from the dataset, the average detection accuracy reaches 39.414%, which is 3.7% higher than YOLOv9, effectively reducing missed detections and improving detection performance.

[0041] To address the problem of low detection accuracy caused by low information richness, the present invention designs a GPR-Trans-Bifpn module that can improve the network capabilities of local area network perception, multi-scale feature fusion, and global correlation across channels and spaces. It can fuse model features of different scales according to weights, enriching the model's information on the detection target features.

[0042] To address the problem of low small target detection accuracy caused by the loss of feature information caused by pooling in traditional downsampling, the present invention also proposes a Coc-ADown module, which uses channel convolution instead of the commonly used pooling downsampling in downsampling, thereby reducing the loss of feature information in downsampling.

[0043] The invention is applicable to the fields of engineering, geological exploration, military affairs, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of the B-scan identification method for small targets of underground facilities of substations by ground penetrating radar as described in the first embodiment;

[0045] Figure 2 This is a schematic diagram of a B-scan image that meets the requirements of the tenth embodiment;

[0046] Figure 3 Schematic diagram of the overall structure of GPR-Bfnet described in Embodiment 10, wherein Upsample represents an upsampling module, Downsample represents a downsampling module, Silence module is a connection layer, Auxiliary Branch represents an auxiliary branch, Head represents a target detection head, Backbone represents the skeleton part of the model, Transformer Connectedlaver represents a fully connected layer connecting the skeleton part of the model, and Neck represents a feature pyramid of the model;

[0047] Figure 4 Schematic diagram of the high- and low-frequency processing branches of GPR-Trans described in Embodiment 10, wherein High-frequency head represents a high-frequency detection head, Low frequency head represents a low-frequency detection head, High-frequencyinformation represents high-frequency information in the image, Normnalization represents a data normalization layer, DWconv represents a depth-wise convolution, Nonlinear activation function represents a nonlinear activation function layer, Maxpool represents maximum pooling, and Softmax represents a nonlinear activation function;

[0048] Figure 5This is a schematic diagram of the Coc-ADown channel convolution down-sampling structure described in Implementation Example 10. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0050] Implementation method 1, see Figure 1 The present embodiment describes a method for identifying small targets of underground facilities of a substation by a ground penetrating radar B-scan, the method comprising:

[0051] S1: Scan the underground facilities along the moving direction according to the UWB ultra-wideband radar system to obtain a B-scan image;

[0052] S2: Label the B-scan image using LabelImage to obtain the characteristic waveform in the B-scan image;

[0053] S3: Build a GPR-Bfnet recognition model and train the GPR-Bfnet recognition model based on the training set;

[0054] S4: Use the trained GPR-Bfnet recognition model to identify and classify the characteristic waveforms in the B-scan image.

[0055] In this embodiment, by constructing the GPR-Bfnet model, the waveform of small targets in the B-scan image can be more accurately identified and classified, reducing false detection and missed detection. The use of LabelImage for automatic identification reduces manual intervention and improves the efficiency and consistency of data processing. The training and application of the GPR-Bfnet model can increase the speed of data processing while maintaining a high degree of accuracy, overcoming the limitations of traditional methods. In view of the specific needs of underground facilities in substations, the model has strong adaptability and pertinence, which improves the reliability of detection.

[0056] Embodiment 2: This embodiment further limits the ground penetrating radar small target B-scan identification method for substation underground facilities described in embodiment 1, and step S1 further includes:

[0057] The B-scan image is enhanced, and new images are generated from the original B-scan image through random cropping, random rotation and mosaic enhancement to construct a data set;

[0058] The data set is divided into training set, test set and validation set according to the ratio of 7:2:1.

[0059] In this embodiment, through image enhancement, random cropping, random rotation and mosaic processing, a variety of new images can be generated, thereby expanding the scale and richness of the data set. This helps to improve the model's ability to recognize small targets of different types and characteristics. Data enhancement can simulate different actual environments and conditions, increase the diversity of the data set, and enable the model to learn more potential variation characteristics during training, thereby improving the performance of the model in practical applications. Dividing the data set into training set, test set and validation set at a ratio of 7:2:1 can ensure that the performance of the model on different data sets is fully evaluated, helping to discover potential overfitting or underfitting problems.

[0060] Implementation method 3: This implementation method further limits the ground penetrating radar small target B-scan recognition method for substation underground facilities described in implementation method 1. The GPR-Bfnet recognition model is constructed in step S3, including:

[0061] Extract low-frequency and high-frequency features through the GPR-Trans module;

[0062] Combining semantic information and location information to process features of different sizes through the GPR-Trans-BiFPN structure;

[0063] Use the bidirectional feature pyramid structure and the Fast Normalized Fusion layer to fuse features of different sizes;

[0064] Use the COC-Adown module for downsampling.

[0065] In this embodiment, the GPR-Trans module extracts low-frequency and high-frequency features in the B-scan image through transformation or convolution operations. Low-frequency features usually correspond to the overall structure or background information of the image, while high-frequency features correspond to the details and edge information of the image. This feature separation helps to better capture details at different levels, thereby improving the accuracy of target recognition. Extracting low-frequency and high-frequency features can help the model understand the multi-level information of the ground penetrating radar image more comprehensively, thereby improving the detection ability of small targets. The GPR-Trans-BiFPN structure combines semantic information and position information, and enhances the detection ability of targets of different scales by processing feature maps of different sizes. It uses a bidirectional information flow to fuse high-level and low-level features, thereby improving the sensitivity of the model to multi-scale targets. This processing method can effectively fuse features at different scales and improve the detection ability of targets of various sizes. The bidirectional feature pyramid structure (BiFPN) optimizes the information flow of feature maps through bottom-up and top-down feature fusion, and combines the fast normalized fusion layer Fast Normalized Fusion for efficient feature fusion. The bidirectional feature pyramid allows the model to capture more contextual information at multiple scales, while FastNormalized Fusion maintains the consistency and effectiveness of features during the fusion process. This fusion method improves the comprehensive utilization of features and helps the model detect and recognize targets more accurately. The COC-Adown module (usually a downsampling module) reduces the spatial resolution in the feature map while maintaining important feature information. The downsampling operation can help reduce the amount of computation and speed up processing while maintaining the model's sensitivity to important features. Through downsampling, the model is able to process image data in a more efficient way, reducing computational complexity while maintaining recognition accuracy.

[0066] In this implementation, the GPR-Trans module can extract low-frequency and high-frequency features, ensuring that the model fully captures image information at different levels, which helps to more accurately identify small targets. The GPR-Trans-BiFPN structure improves the recognition ability of targets of different scales by processing and fusing features of different sizes, especially for small targets in underground facilities of substations. The bidirectional feature pyramid structure and Fast Normalized Fusion can effectively integrate multi-scale features, improve the quality and efficiency of feature fusion, and enhance the model's adaptability to complex scenes. The downsampling operation of the COC-Adown module reduces the computational complexity while maintaining the necessary feature information, so that the model maintains a high recognition accuracy while operating efficiently. Through the comprehensive application of these modules, the GPR-Bfnet model can handle various complex situations and improve the robustness and overall performance of small target recognition by ground penetrating radar.

[0067] Implementation method 4: This implementation method further limits the B-scan recognition method for small targets of underground facilities of substations by ground penetrating radar described in implementation method 3. The GPR-Trans module includes a dual-branch information processing architecture, and uses a maximum pooling layer after the normalization layer to process low-frequency information:

[0068] X global =Attention(Q g ,Pool(K g ),Pool(V g ))

[0069] Among them, Qg represents the global query vector, Kg represents the global key vector, Vg represents the global value vector, and Xglobal represents the vector of the global branch;

[0070] Use depth-wise separable convolution DWConv to process high-frequency information;

[0071] Q,K,V=FC(X in )

[0072] V s =DWconv(V)

[0073] Q l =DWconv(Q)

[0074] K l =DWconv(K)

[0075]

[0076] X local =Attn⊙V s

[0077] Among them, Q l represents the local branch query vector, K l represents the local branch key vector, V s represents a local numerical vector, X in represents the original input vector, Q represents the query vector, K represents the key vector, V represents the numerical vector, DWconv represents the depth-wise convolution, A ttnt represents the attention mechanism, A ttn Indicates the addition of nonlinear activation function attention mechanism, X local Represents local branch output.

[0078] This implementation uses a dual-branch information processing architecture to process low-frequency information and high-frequency information separately. The core idea of ​​this architecture is to use different processing methods to extract and optimize information in different frequency bands to improve recognition accuracy.

[0079] Normalize low-frequency information to ensure the consistency of input data, reduce the difference in numerical range, and improve the stability and convergence speed of model training. After the normalization layer, a maximum pooling layer is used to extract significant features from low-frequency information. The maximum pooling operation can reduce the data dimension while retaining important feature information, which is helpful for subsequent feature processing and target recognition. High-frequency information is processed by depthwise separable convolution. Depthwise separable convolution decomposes the standard convolution operation into channel-by-channel convolution and point-by-point convolution, which can significantly reduce the amount of calculation and the number of parameters while maintaining efficient feature extraction capabilities.

[0080] The global query vector (Qg), global key vector (Kg) and global value vector (Vg) are used to process global feature information. The global branch helps to capture the overall context information and integrates the global information through the vector representation (Xglobal).

[0081] The local branch query vector (Ql) and the local branch key vector (Kl) are used to process local feature information. The local branch focuses on the details of local features, extracts local information through the original input vector (Xin), and enhances the recognition ability of small targets. The dual-branch architecture allows the model to learn features at both the global and local levels. The global branch can process the overall context information, while the local branch focuses on details and local features. This dual processing method makes the feature expression more comprehensive and rich.

[0082] Embodiment 5: This embodiment further limits the ground penetrating radar small target B-scan recognition method for underground facilities of substations described in embodiment 3. The bidirectional feature pyramid structure and the fast normalized fusion layer FastNormalized Fusion perform feature fusion of different sizes, including:

[0083] Establish two pathways, top-down and bottom-up; the top-down pathway is used for feature fusion, and the bottom-up pathway is used for transmitting position information;

[0084] Depthwise separable convolution is used for feature fusion, and batch normalization and activation are added after each convolution.

[0085] This implementation method optimizes the B-scan recognition of small targets of ground penetrating radar in underground facilities of substations by establishing a bidirectional feature pyramid structure (top-down and bottom-up pathways) and applying Fast Normalized Fusion. The principle of this design is to use the top-down pathway for feature fusion to combine high-level semantic information with low-level detail information; the bottom-up pathway is responsible for accurately transmitting location information, thereby improving the accuracy of target detection. Depth-wise separable convolution further enhances the feature extraction capability, while combining batch normalization and activation functions to improve the stability and performance of the network. The advantage of this method is that it can better handle multi-scale feature fusion, improve the recognition ability of small targets of different sizes, and due to the use of batch normalization and activation, the training efficiency and accuracy of the model are also improved.

[0086] Implementation method 6. This implementation method is a further limitation of the ground penetrating radar small target B-scan recognition method for underground facilities of substations described in implementation method 3. The COC-Adown channel stacking convolution downsampling replaces the traditional pooling layer, divides the image into sub-scales, and downsamples by channel stacking, and finally uses a single-step convolution to optimize the amount of calculation, specifically:

[0087] f 0,0 =X[0:S:scale,0:S:scale],

[0088] f 1,0 =X[1:S:scale,0:S:scale],...,

[0089] f scale-1,0 =X[scale-1:S:scale,0:S:scale];

[0090] f 0,1 =X[0:S:scale,1:S:scale],f 1,1 ,...,

[0091] f scale-1,1 =X[scale-1:S:scale,1:S:scale];

[0092] f scale-1 =X[0:S:scale,scale-1:S:scale],f 1,scale-1 ,...

[0093] f scale-1,scale-1 =X[scale-1:S:scale,scale-1:S:scale]

[0094] Among them, X represents the feature map, fx,y represents a sub-image, S represents the size of the input feature map, and Scale represents the multiple of the spatial dimension change.

[0095] This implementation replaces the traditional pooling layer with the COC-Adown channel stacking convolution, divides the image into sub-scales and performs downsampling, thereby optimizing the feature extraction process. Specifically, when downsampling, the COC-Adown channel stacking convolution reduces the amount of calculation and improves the ability to retain features through the stacking convolution operation. Compared with the traditional pooling layer, this method can reduce the computational complexity while retaining more detail information. The advantages include improving the expressiveness of the feature map and the flexibility of downsampling, while avoiding the loss of details that may be caused by pooling.

[0096] Embodiment 7: The ground penetrating radar small target B-scan recognition system for underground facilities of substations described in this embodiment includes:

[0097] A scanning unit, used for scanning the underground facilities along the moving direction according to the UWB ultra-wideband radar system to obtain a B-scan map;

[0098] A characteristic waveform acquisition unit, used to label the B-scan image using LabelImage, and acquire a characteristic waveform in the B-scan image;

[0099] A model building unit, used for building a GPR-Bfnet recognition model and training the GPR-Bfnet recognition model according to a training set;

[0100] The recognition unit is used to recognize and classify the characteristic waveform in the B-scan image using the trained GPR-Bfnet recognition model.

[0101] Embodiment 8. A computer device described in this embodiment includes a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the ground penetrating radar small target B-scan identification method for underground facilities of a substation according to any one of embodiments 1 to 6.

[0102] Embodiment 9. A computer-readable storage medium described in this embodiment stores a computer program, which, when executed by a processor, executes the steps of the ground-penetrating radar small target B-scan identification method for underground facilities of a substation as described in any one of claims 1 to 7.

[0103] Implementation Method 10: See Figures 2 to 5This embodiment provides a specific example of the ground penetrating radar small target B-scan identification method for the substation underground facilities described in the first embodiment, and is also used to explain the second to sixth embodiments, specifically:

[0104] Firstly, the UWB ultra-wideband radar system is used to scan the underground facilities with complex structures along the direction of movement to obtain the B-scan image. The features are marked with labelimg and then put into the GPR-Bfnet model. The weights are set for training. The reasonableness and reliability of the training weights are judged by observing the test set and the validation set and combining the confidence level.

[0105] The framework aims to develop a deep learning-based ground penetrating radar data processing system to improve the accuracy of small target sample recognition in the ground penetrating radar B-scan for underground facilities in substations.

[0106] Specifically, it is divided into the following steps:

[0107] (1) Pretreatment

[0108] Before field detection, ground survey and site investigation are carried out to understand the distribution of underground facilities to be detected. Then, the UWB ultra-wideband radar system is used to perform a plane push scan of the ground to be detected to obtain a B-scan image that meets the requirements. Figure 2 As shown in the figure, multiple B-scan images are enhanced, and new images are generated from the basic data set by random cropping, random rotation and mosaic enhancement, and organized into data sets, which are divided into training set, test set and validation set according to 7:2:1. Each characteristic waveform in the data set is classified and labeled in labalimg.

[0109] (2) Training

[0110] In order to achieve better detection results for targets with different pixels in the B-scan dataset of the ground penetrating radar and improve the missed detection of small targets that account for a large proportion of the dataset, the structure of YOLOv9 is improved. Figure 3 As shown in the figure, similar to the programmable gradient information, since the information loss in the process of pooling and information transmission is the main reason for the high missed detection rate of small target recognition, enriching the incoming information and reducing the pooling steps are one of the keys to reduce the missed detection rate of small targets. Therefore, a miniaturized transformer architecture is used in this embodiment to enrich the feature extraction.

[0111] The dual-branch architecture attention mechanism is used to locally enhance the backbone network, and the improved network structure is shown in the figure. Compared with YOLOv9, the dual-branch attention mechanism is used in this embodiment to enhance the backbone network, and the convolution part of the original backbone is assembled into a GPR-Trans model to reduce the number of times the RepNCSPELAN4 module is reused, and efficient feature extraction is achieved. The features of different scales output after GPR-Trans processing are placed in a pyramid architecture that is highly consistent with the scale gradient, and a bidirectional feature pyramid structure is designed. A bidirectional feature fusion mechanism and a weighted feature fusion method are introduced to overcome the problems of low feature fusion efficiency and insufficient information flow. Weights are introduced in the feature fusion stage to effectively integrate features at all levels, adjust the transmission of information to bidirectional, and make this part of the structure form an effective stack, so as to generate more discriminative features for subsequent structures. Since the attention-guided bidirectional feature pyramid network has a more efficient feature generation capability, the RepNCSPELAN4 module is reduced in this embodiment to obtain a smaller network structure as much as possible. The structural framework and principle of each module will be introduced below. Based on the analysis of information loss by YOLOv9, this implementation method links information of different sizes multiple times in the model. In order to facilitate the exchange of information, this implementation method adopts a miniaturized structure of transformer, so that detection targets of different sizes can be effectively detected. Therefore, this implementation method proposes GPR-Transformer. Figure 4 shown.

[0112] GPR-Trans not only solves the problem of transmitting and combining features of different sizes, but also considers the capture of low-frequency information and high-frequency information. Faced with the problem that the same branch is difficult to simultaneously meet the effective detection of low-frequency information and high-frequency information, GPR-Trans proposes a dual-branch information processing architecture. After the normalization layer, the maximum pooling layer is used to process low-frequency information to reduce the amount of computation of low-frequency information that accounts for a large proportion. The input data interacts with the original matrix multiple times and cooperates with the nonlinear activation function to capture the global low-frequency information. The formula is as follows:

[0113] X global =Attention(Q g ,Pool(K g ),Pool(V g ))

[0114] Among them, Qg represents the global query vector, Kg represents the global key vector, Vg represents the global value vector, and Xglobal represents the vector of the global branch;

[0115] The high-frequency information branch uses deep separable convolution to reduce the amount of computation of the model and generate shared weights. It is worth noting that in order to retain the small target information of the high-frequency branch, this implementation abandons the inherent pooling method to reduce the amount of computation, and uses DWConv to retain as much information as possible from the original image. In this implementation, DWConv is also introduced twice in this model block to generate context-aware weights for multiple feature fusions, and finally a nonlinear activation function is introduced to give the attention mechanism stronger nonlinearity. The formula is as follows:

[0116] Q,K,V=FC(X in )

[0117] V s =DWconv(V)

[0118] Q l =DWconv(Q)

[0119] K l =DWconv(K)

[0120]

[0121] X local =Attn⊙V s

[0122] Among them, Q l represents the local branch query vector, K l represents the local branch key vector, V s represents a local numerical vector, X in Represents the original input vector. First, two DW-convs are used to aggregate Q l and K l Then, to obtain a context-aware weight between -1 and 1, l and K l A series of transformations are performed by performing the Hadamard product of . Finally, this weight is used to further enhance the local features.

[0123] The targets in the B-scan image of the ground penetrating radar are mostly high-frequency information parts. Therefore, the addition of high-frequency branches and low-frequency branches can, to a certain extent, specialize the detection effect of small targets while retaining the original detection effect of larger targets, avoiding the feature loss caused by pooling, and optimizing the detection effect of small and medium targets by ground penetrating radar.

[0124] In the Neck part of GPR-YOLOv9, in order to better cooperate with the output of the backbone GPR-Trans layer, the GPRtrans-bifpn structure is proposed in this embodiment to process and combine outputs of different sizes. The processed five different size features are placed in five layers, and then through the bipfn structure, two top-down and bottom-up paths are established. Among them, the top-down path is used for feature fusion, which can make the feature map used for prediction have higher semantic information after fusion, thereby improving the prediction accuracy. The bottom-up path is responsible for transmitting the position information to the predicted feature map, so that the predicted feature map has both higher semantic information and position information. Finally, in order to fuse feature information of different sizes, Fast normalized fusion is introduced to fuse the features. It is worth noting that in order to further improve efficiency, depth-wise separable convolution is used for feature fusion, and batch normalization and activation are added after each convolution, which is conducive to collecting feature information at different scales, and finally enriching the information degree of features in the model, which is conducive to improving detection accuracy.

[0125] In the downsampling part of GPR-YOLOv9, in order to retain more information of the original features and avoid the loss of key small target features in the pooling during downsampling, we combined the original lightweight and flexible downsampling module ADown with Channeloverlay convolution. The purpose is to reduce the loss of small target information at the cost of increasing the amount of small operations. The introduction of COC-Adown is consistent with the idea of ​​reducing information loss in this article. In order to ensure that the rich information in the above text is lost as little as possible during downsampling, we use COC to replace the original ADown pooling layer. The principle of the combined downsampling structure and COC layer is as follows: Figure 5 shown.

[0126] The essence of the COC layer is to divide the image into divisible scales, downsample the sub-scales by channel superposition, reduce the size of the image, enrich the information channel, and finally use single-step convolution to compress the channel to optimize the problem of increased computation. The purpose of adding this module to the model is to replace the pooling part in the traditional downsampling to reduce the loss of feature information in the pooling process and provide more feature information for the model to detect small objects.

[0127] Combining the above two models, the designed detection model takes enriching the feature information of small targets as the entry point, proposes the GPR-Trans-Bifpn module to process high-frequency information and fuse feature information of different sizes in the model, and designs the Coc-ADown module to replace the pooling process to reduce the loss of small target feature information caused by downsampling.

[0128] Those skilled in the art will appreciate that the embodiments of the present disclosure may be provided as methods, systems or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] The present disclosure is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram and the combination of processes and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0130] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present disclosure rather than to limit its protection scope. Although the present disclosure has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that after reading the present disclosure, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the disclosed claims to be approved.

Claims

1. A B-scan recognition method for small targets of ground penetrating radar for underground facilities of substations, characterized in that: The method comprises: S1: Scan the underground facilities along the moving direction according to the UWB ultra-wideband radar system to obtain a B-scan image; S2: Label the B-scan image using LabelImage to obtain the characteristic waveform in the B-scan image; S3: Build a GPR-Bfnet recognition model and train the GPR-Bfnet recognition model based on the training set; S4: Use the trained GPR-Bfnet recognition model to identify and classify the characteristic waveforms in the B-scan image.

2. The method for identifying small targets of underground facilities of substation by ground penetrating radar B-scan according to claim 1 is characterized in that: The step S1 further comprises: The B-scan image is enhanced, and new images are generated from the original B-scan image through random cropping, random rotation and mosaic enhancement to construct a data set; The data set is divided into training set, test set and validation set according to the ratio of 7:2:

1.

3. The method for identifying small targets of underground facilities of substation by ground penetrating radar B-scan according to claim 1 is characterized in that: The step S3 constructs a GPR-Bfnet recognition model, including: Extract low-frequency and high-frequency features through the GPR-Trans module; Combining semantic information and location information to process features of different sizes through the GPR-Trans-BiFPN structure; Use the bidirectional feature pyramid structure and the Fast Normalized Fusion layer to fuse features of different sizes; Use the COC-Adown module for downsampling.

4. The method for identifying small targets of underground facilities of substation by ground penetrating radar B-scan according to claim 3 is characterized in that: The GPR-Trans module includes a dual-branch information processing architecture, using a maximum pooling layer after the normalization layer to process low-frequency information: X global =Attention(Q g ,Pool(K g ),Pool(V g )) Among them, Qg represents the global query vector, Kg represents the global key vector, Vg represents the global value vector, and Xglobal represents the vector of the global branch; Use depth-wise separable convolution DWConv to process high-frequency information; Q,K,V=FC(X in ) V s =DWconv(V) Q l =DWconv(Q) K l =DWconv(K) X local =Attn⊙V s Among them, Q l represents the local branch query vector, K l represents the local branch key vector, V s represents a local numerical vector, X in represents the original input vector, Q represents the query vector, K represents the key vector, V represents the numerical vector, and DWconv represents the depth-wise convolution. represents the attention mechanism, A ttn Indicates the addition of nonlinear activation function attention mechanism, X local Represents local branch output.

5. The method for identifying small targets of underground facilities of substation by ground penetrating radar B-scan according to claim 3 is characterized in that: The bidirectional feature pyramid structure and the fast normalized fusion layer Fast Normalized Fusion perform feature fusion of different sizes, including: Establish two pathways, top-down and bottom-up; the top-down pathway is used for feature fusion, and the bottom-up pathway is used for transmitting position information; Depthwise separable convolution is used for feature fusion, and batch normalization and activation are added after each convolution.

6. The method for identifying small targets of ground penetrating radar B-scan for underground facilities of substation according to claim 3 is characterized in that: The COC-Adown channel stacking convolution downsampling replaces the traditional pooling layer, divides the image into sub-scales, and downsamples by channel stacking, and finally uses single-step convolution to optimize the amount of calculation.

7. The ground penetrating radar small target B-scan recognition system for underground facilities of substations is characterized by: The system comprises: A scanning unit, used for scanning the underground facilities along the moving direction according to the UWB ultra-wideband radar system to obtain a B-scan map; A characteristic waveform acquisition unit, used to label the B-scan image using LabelImage, and acquire a characteristic waveform in the B-scan image; A model building unit, used for building a GPR-Bfnet recognition model and training the GPR-Bfnet recognition model according to a training set; The recognition unit is used to recognize and classify the characteristic waveform in the B-scan image using the trained GPR-Bfnet recognition model.

8. A computer device, characterized in that: It includes a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the ground penetrating radar small target B-scan identification method for underground facilities of a substation according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the steps of the ground-penetrating radar small target B-scan identification method for underground facilities of a substation as described in any one of claims 1 to 6.

Images