Reinforcing steel bar GPR signal identification method based on hyperbolic attention mechanism

Abstract

The invention relates to the technical field of detection, in particular to a reinforcing steel bar GPR signal recognition method based on a hyperbolic attention mechanism, and the method comprises the following steps: S1, signal preprocessing: carrying out the DC offset removal and noise reduction of radar reflection signal data obtained by a ground penetrating radar; s2, feature extraction: hyperbolic features in the radar image are extracted through a ResNet50 network including a depth separable convolution and a deformable convolution; s3, regional proposal: generating a candidate region containing hyperbolic characteristics by using a regional proposal network; s4, hyperbolic attention optimization: enhancing the positioning precision of hyperbolic features in the candidate region through a hyperbolic attention mechanism (HAT); and S5, target positioning and analysis: carrying out recognition and distribution analysis on the positions of the reinforcing steel bars based on the optimized features. According to the method, the DAS module and the HAT module can be utilized, and the model is trained by combining GprMax synthesis and on-site acquired images, so that accurate recognition of the reinforcing steel bars in the B-Scan image of the tunnel lining GPR is realized.

Description

Technical Field

[0001] The present invention relates to the field of detection technology, and particularly to a method for identifying steel bar GPR signals based on a hyperbolic attention mechanism. Background Art

[0002] High-reflectivity substances such as steel bars and metal pipes have much greater blocking and reflection capabilities for electromagnetic waves than transmission capabilities. Their existence has important reference value for building safety monitoring and underground pipeline detection. In GPR images, underground targets such as steel bars usually present a hyperbolic shape. This is because when the electromagnetic waves emitted by the ground-penetrating radar encounter underground targets, they will be reflected back to the ground-penetrating radar receiver at different time points. The multiple reflections of the radar waves in the steel bar and its surrounding environment result in multiple hyperbolas appearing in the radar image. These hyperbolas represent the reflection results after the radar waves propagate through different paths. For the hyperbolas presented by the steel bars in the radar image, through convolution operations, local features such as edges and corners in the image can be extracted. In the ground-penetrating radar image, the hyperbolic features of the steel bars are exactly a manifestation form of these local features, that is, the change of amplitude. In the non-destructive testing of ground-penetrating radar, the traditional method is to manually interpret the image through the hyperbolic features in the output display radar grayscale image or use data processing to identify the steel bars through signal feature analysis, while the currently popular method is to use deep learning to extract the features in the radar grayscale image to achieve the detection of targets in the tunnel lining, thus improving the detection efficiency. Summary of the Invention

[0003] Aiming at the technical problems existing in the background art, the present invention provides a method for identifying steel bar GPR signals based on a hyperbolic attention mechanism, which uses DAS and HAT modules, combines GprMax synthesis and field-collected image training models, and realizes the accurate identification of steel bars in the B-Scan image of tunnel lining GPR.

[0004] The technical implementation scheme of the present invention is as follows:

[0005] A method for identifying steel bar GPR signals based on a hyperbolic attention mechanism, comprising the following steps:

[0006] S1. Signal preprocessing: removing the DC offset and reducing the noise of the radar reflection signal data obtained by the ground-penetrating radar;

[0007] S2. Feature extraction: extracting the hyperbolic features in the radar image through a ResNet50 network including depthwise separable convolution and deformable convolution;

[0008] S3. Region proposal: using a region proposal network to generate candidate regions containing hyperbolic features;

[0009] S4. Hyperbolic Attention Optimization: Enhance the localization accuracy of hyperbolic features in candidate regions through the Hyperbolic Attention Mechanism (HAT);

[0010] S5. Target Localization and Analysis: Identify the positions of steel bars and analyze their distribution based on the optimized features.

[0011] Optionally, the signal preprocessing includes:

[0012] A. Remove the DC offset from the single-channel reflected wave data to produce the processed amplitude data;

[0013] B. Construct a B-Scan image by continuously scanning the data to present the hyperbolic features.

[0014] Optionally, the feature extraction includes:

[0015] A. Deformable convolution is used to dynamically adjust the receptive field to adapt to the hyperbolic shape;

[0016] B. Depthwise separable convolution realizes lightweight feature extraction.

[0017] Optionally, the Hyperbolic Attention Mechanism (HAT) includes:

[0018] A. Calculate the hyperbolic distance between feature points based on the hyperbolic geometric model;

[0019] B. Enhance the hyperbolic feature response through attention weight assignment.

[0020] Optionally, it also includes generating simulated radar images using GprMax and training the model by combining simulated data with real data.

[0021] Optionally, the Region Proposal Network generates candidate regions through multi-scale anchor boxes and filters out valid proposals through non-maximum suppression.

[0022] Optionally, the target localization and analysis steps include:

[0023] A. Perform ROI pooling and feature alignment on the candidate regions;

[0024] B. Achieve precise determination of the positions of steel bars through a classifier and a regressor.

[0025] Optionally, the method is applicable to the non-destructive testing of steel bars in tunnel lining structures, and can identify multiple hyperbolic reflection features and analyze their distribution patterns.

[0026] The present invention has the following advantages:

[0027] This method uses a deep learning model focused on hyperbolic characteristics, enhances the attention to hyperbolas with the deformable attention mechanism DAS, adds a hyperbolic attention HAT module to strengthen the recognition of the hyperbolic characteristics of steel bars, and trains the model with GprMax synthetic images and field-collected images to realize the recognition of steel bars in tunnel lining radar images, providing a new approach for the recognition of tunnel structure steel bars. Description of the Drawings

[0028] Figure 1 It is a model diagram for GPR to recognize steel bars;

[0029] Figure 2 It is a schematic diagram of the hyperbolic formation principle of steel bars in GPR images;

[0030] Figure 3 It is a ResNet50 diagram integrated with the deformable attention mechanism (A Deformable Attention to Capture Salient Information, DAS);

[0031] Figure 4 It is a diagram of the hyperbolic attention (Hyperbolic Attention, HAT) module;

[0032] Figure 5 It is a diagram of the difference in steel bar recognition between the front and back models. Detailed Implementation Manner

[0033] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. It is hereby declared that the orientation terms such as up, down, left, right, front, back, inside, and outside that appear or will appear in the text of the present invention are only based on the accompanying drawings of the present invention, and they do not specifically limit the present invention.

[0034] Example 1:

[0035] A method for identifying GPR signals of steel bars based on the hyperbolic attention mechanism includes the following steps:

[0036] S1. A signal processing module is used to preprocess the tunnel lining radar reflection signal data obtained by a ground penetrating radar, including DC offset removal and noise reduction; S2. A feature extraction module extracts the hyperbolic features of steel bars in a radar image based on a ResNet50 model that includes depthwise separable convolution (DSC) and deformable convolution (DC); S3. A region proposal module generates candidate regions through a Region Proposal Network (RPN) to identify hyperbolic morphological features that may contain steel bars; S4. A hyperbolic attention recognition module optimizes the positioning accuracy of features in candidate regions through a hyperbolic attention mechanism (HAT); S5. A positioning and analysis module accurately identifies and locates the identified steel bars to determine whether they meet the design specifications.

[0037] It should be noted that in step S1, the signal data obtained by the ground penetrating radar includes A-Scan single-channel scan data and B-Scan continuous scan data. Among them, the A-Scan data records the variation of the amplitude of a single-channel radar signal over time, and the B-Scan data is generated by continuous scanning and is used to construct the hyperbolic features of steel bars. A DC offset removal operation is performed on each reflected wave. By processing the amplitude of the sampling points, the amplitude data after DC offset removal is obtained, and the processed reflected wave image is regenerated to improve the accuracy of the signal and reduce the influence of DC offset on the signal.

[0038] It should be noted that in step S4, by jointly training the radar data with simulation data and real data, and using the finite-difference time-domain technique to generate simulation images, the physical characteristics of ground penetrating radar images can be accurately captured. This method enhances the generalization ability of the model when detecting steel bars in real ground penetrating radar images, especially under various interference conditions.

[0039] It should be noted that in the above steps, the hyperbolic distance of the reflection signal points is calculated through a hyperbolic geometric model to further optimize the extraction and discrimination ability of the hyperbolic features of steel bars and ensure the accuracy of the signal features; by introducing a hyperbolic attention mechanism (HAT), the recognition ability of the model for the multiple hyperbolic reflection features of steel bars in radar images is enhanced to ensure the accuracy of the position and distribution of steel bars. Through this method, the distribution of steel bars in the tunnel lining can be accurately identified. This method improves the recognizability of the data and ensures the accurate identification of steel bars in the tunnel lining.

[0040] Embodiment 2:

[0041] The specific operation method is as follows:

[0042] Step 1. Layout of survey lines and measuring points and data collection

[0043] A survey line is carefully arranged at predetermined intervals on the tunnel lining surface, and m measuring points are set in order. For the i-th measuring point (i=1,2,...,m), the corresponding horizontal position coordinate is xi, and the ground penetrating radar (GPR) technology is used to detect the steel bars at each measuring point. The collected .dzt data usually contains multiple data channels, each set of data corresponds to a detection position, that is, an A-scan, and many single-channel data are cleverly combined to generate the corresponding B-scan scanning image. In this image, the steel bars present a hyperbolic shape, and the image completely contains the amplitude and phase directional information of the reflected signal, providing a rich initial data foundation for subsequent in-depth analysis.

[0044] In this process, the change in signal strength can be described by observing the amplitude of the signal in the ground penetrating radar data. When the electromagnetic wave enters the concrete steel bar, when the electromagnetic wave propagates to two layers of different media, the amplitude A can be divided into two directions: positive phase and negative phase. The direction of the positive phase is the same as the phase direction of the incident wave, indicating that the amplitude A maintains a consistent phase relationship with the incident wave. The main reason for the amplitude change is the change in the positive and negative values of the reflection coefficient. When the electromagnetic wave enters the steel bar from the concrete, the Fresnel reflection coefficient is negative, causing a phase change. The phase change information of these reflected waves provides rich data for subsequent analysis, helping to further infer the characteristics of the steel bar signal.

[0045] Step 2: Data Preprocessing

[0046] (I) Accurate removal of DC offset: In the i-th reflected wave, n sampling points are strictly selected at equal intervals along the direction of gradually increasing two-way travel time t. For each selected sampling point j (j = 1, 2, ..., n), the new amplitude A1 after removing the DC offset is accurately calculated. The original amplitudes of these n sampling points are processed in turn using this calculation method to reconstruct the processed i-th reflected wave. After this processing, the interference of DC offset on the signal is effectively reduced, and the degree of signal distortion is significantly reduced, providing more accurate and reliable data support for subsequent in-depth analysis.

[0047] (II) Simulation data expansion and enhancement: After completing the preliminary processing of the original signal, the simulated image data carefully generated by GprMax using the finite difference method are integrated into the formed image data set. These simulated data can highly accurately display the physical characteristics of the ground penetrating radar image. By adding simulated data, the model's generalization ability to detect steel bars when facing real ground penetrating radar images (especially images with various complex interference conditions) is greatly enhanced, effectively improving the model's adaptability and reliability.

[0048] Step 3: Multi-dimensional feature extraction and recognition

[0049] (1) Deep feature mining of convolutional neural network: The images in the dataset are input into a carefully constructed convolutional neural network one by one. Through multiple convolutional operations inside the network, feature maps are deeply extracted. Each pixel in the feature map corresponds to a specific area in the original image, that is, the receptive field. The feature map not only completely retains the spatial structure information of the original image but also deeply refines the high-level semantic information, laying a solid feature foundation for subsequent analysis and recognition.

[0050] (2) Focus on the key with hyperbolic space attention mechanism: Compared with the traditional Euclidean space, the hyperbolic space has a more powerful non-linear feature representation ability. In the traditional Euclidean space, for complex geometric shapes, it often requires building highly complex models or performing a large number of feature combinations to achieve a relatively accurate representation. However, in the hyperbolic space, some complex structures can be represented in a more natural, compact, and concise way, which greatly improves the efficiency of the model in capturing complex shapes.

[0051] By introducing the attention mechanism of the hyperbolic space, the model can more efficiently focus its attention on the key structures and important information in the image. For example, for the area where the steel bar hyperbola is located, the attention of the model will significantly tend to such key features, greatly enhancing the importance of the key features and effectively reducing the attention to noise or irrelevant features. The attention mechanism calculated using the hyperbolic distance has outstanding advantages in recognizing complex forms such as curves or non-linear boundaries and can significantly improve the model's recognition ability for such complex image structures.

[0052] (3) Accurately screen candidates with the Region Proposal Network: On the obtained feature map, the Region Proposal Network (RPN) in Faster R-CNN is introduced. The core task of the RPN is to accurately generate candidate regions, that is, those regions that may contain the target (steel bar). The RPN generates multiple anchor boxes with different sizes and aspect ratios by sliding a window on the feature map. After a series of calculations and screenings, the RPN outputs a group of carefully screened region proposals (bounding boxes), which provide a very reliable basis for the subsequent steel bar detection work and effectively ensure the accuracy and stability of the detection results.

[0053] (4) ROI Pooling and Precise Result Output: The candidate boxes generated by RPN are accurately mapped to the feature map, and the features of the corresponding regions are extracted. Since the sizes of these feature regions are different, in order to facilitate subsequent classification and regression operations, the ROI pooling technology is used to unify the sizes of the feature regions; the features after ROI pooling processing are respectively sent into two parallel branches, one branch is used for object classification, and the other branch is used for bounding box regression. Finally, the model outputs the accurate class label, confidence score, and the accurately adjusted bounding box coordinates of each candidate region.

[0054] Through the above series of rigorous and efficient steps, this solution can accurately and quickly identify the steel bar distribution in tunnel lining steel bar detection. Even in the face of interference images that the model has never learned, with the powerful advantages of the hyperbolic attention mechanism, it can still reliably ensure the accurate identification of steel bars in the tunnel lining, providing strong support for the safety assessment of the tunnel lining structure.

[0055] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments, and various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art.

Claims

1. A method for identifying steel bar GPR signals based on a hyperbolic attention mechanism, characterized in that It includes the following steps: S1. Signal preprocessing: removing the DC offset and reducing the noise from the radar reflection signal data obtained by the ground penetrating radar; S2. Feature extraction: extracting the hyperbola features in the radar image through the ResNet50 network including depthwise separable convolution and deformable convolution; S3. Region proposal: generating candidate regions containing hyperbola features by using the region proposal network; S4. Hyperbolic attention optimization: enhancing the localization accuracy of the hyperbola features in the candidate regions through the hyperbolic attention mechanism (HAT); S5. Target localization and analysis: identifying and analyzing the position and distribution of steel bars based on the optimized features.

2. The method for identifying steel bar GPR signals based on a hyperbolic attention mechanism according to claim 1, wherein The signal preprocessing includes: A. Removing the DC offset from the single-channel reflection wave data to produce the processed amplitude data; B. Constructing a B-Scan image by continuously scanning the data to present the hyperbola features.

3. The steel bar GPR signal recognition method based on a hyperbolic attention mechanism according to claim 1, characterized in that, The feature extraction includes: A. Deformable convolution is used to dynamically adjust the receptive field to adapt to the hyperbola shape; B. Depthwise separable convolution realizes lightweight feature extraction.

4. A method for identifying steel bar GPR signals based on a hyperbolic attention mechanism according to claim 1, characterized in that, The hyperbolic attention mechanism (HAT) includes: A. Calculating the hyperbolic distance between feature points based on the hyperbolic geometry model; B. Enhancing the hyperbola feature response through attention weight assignment.

5. A method for identifying steel bar GPR signals based on a hyperbolic attention mechanism according to claim 1, characterized in that, It also includes generating simulated radar images by using GprMax and training the model by combining the simulated data with the real data.

6. A method for identifying steel bar GPR signals based on a hyperbolic attention mechanism according to claim 1, characterized in that, The region proposal network generates candidate regions through multi-scale anchor boxes and filters the effective proposals through non-maximum suppression.

7. A method for identifying steel bar GPR signals based on a hyperbolic attention mechanism according to claim 1, characterized in that, The target localization and analysis step includes: A. Performing ROI pooling and feature alignment on the candidate regions; B. Achieving the accurate determination of the steel bar position through the classifier and the regressor.

8. According to the method described in any one of claims 1-7, characterized in that, The method is applicable to the non-destructive detection of steel bars in the tunnel lining structure, and can identify multiple hyperbola reflection features and analyze their distribution rules.

Images