Automatic detection method of hidden cracks based on multi-feature fusion of radar signals

Abstract

The present invention discloses an automatic detection method for hidden cracks based on multi-feature fusion of radar signals, comprising the following steps: step 1, collecting actual highway sections to obtain ground penetrating radar raw data; step 2, processing ground penetrating radar signal data and constructing a multi-feature ground penetrating radar data set; step 3, designing a two-stage through crack detection model based on Swin Transformer optimization; step 4, designing a one-stage through crack detection model based on Swin Transformer-YOLOv8 optimization. The present invention can identify and locate road cracks more quickly and accurately, reduce maintenance time and cost, reduce safety risks such as road collapse caused by cracks, enhance the model's ability to identify different types of cracks, improve the generalization of the detection model, help maintain the integrity of the road and extend the service life of the road, can continuously monitor road conditions, provide crack detection data in real time, provide a scientific basis for road maintenance decisions, and improve road safety.

Description

Technical Field

[0001] The present invention relates to the fields of road disease detection and intelligent image recognition, and in particular to a method for automatically detecting hidden cracks based on multi-feature fusion of radar signals. Background Art

[0002] Road defect detection is a crucial component of road maintenance and management. With the increasing demand for transportation, the impact of road surface quality on traffic safety and efficiency is becoming increasingly significant. Traditional road defect detection methods primarily rely on manual inspections and simple instrument testing, which are inefficient and susceptible to subjective factors, resulting in inaccurate and unreliable detection results. Furthermore, these methods often require significant human and material resources, making them difficult to meet the real-time and precision requirements of modern road management. Therefore, the development of efficient and intelligent detection methods is an urgent issue. Ground-penetrating radar (GPR) technology has garnered widespread attention due to its ability to non-invasively obtain information about the internal structure of roads. By emitting electromagnetic waves and receiving reflected signals, GPR can deeply examine the various layers of a road's structure, providing high-resolution subsurface information in real time. This technology shows promising application prospects in detecting potential defects and assessing road health.

[0003] Meanwhile, with the rapid development of artificial intelligence (AI), deep learning-based object detection has made significant progress in computer vision. These technologies not only improve image processing efficiency but also enable highly accurate object detection in complex scenes. Combining ground-penetrating radar (GPR) technology with deep learning-based object detection can significantly improve the efficiency and accuracy of road defect detection, providing a scientific basis for road maintenance.

[0004] Existing crack detection methods based on ground-penetrating radar are mainly divided into several types. One type is the convolutional neural network model based on deep learning, another type is the feature extraction model based on traditional machine learning, and another type is the technical solution based on image segmentation. In addition, the Transformer-based detection model has gradually emerged in recent years. In addition, existing research mainly relies on convolutional neural networks. Although certain results have been achieved, its ability to capture global information is limited.

[0005] The shortcomings of the above methods in the process of road crack detection mainly include: (1) There is a relative lack of research on the recognition and detection of crack penetration, which leads to limited means of evaluating the safety and stability of road structures; (2) Most of the existing target detection models are based on convolutional neural networks (CNNs), which have limitations in global information capture and correlation understanding. Especially when dealing with cracks with complex shapes, it is difficult to accurately capture the overall structure and context information of the cracks, thus affecting the accuracy and completeness of the detection results; (3) Due to the limitations of detection equipment and the complexity of the underground environment, the current detection technology leads to a limited range of recorded signal amplitude intensity. When mapped to a grayscale image, the pixel range is limited, and the image quality is degraded due to signal attenuation, which affects the accurate capture of crack features and causes misjudgment and missed judgment; (4) The existing methods do not fully utilize time-frequency domain information in the signal processing process, which limits the multi-angle analysis and representation of crack features, resulting in poor extraction and recognition of crack features.

[0006] This method leverages non-destructive detection data from ground-penetrating radar equipment on road cracks, combined with a deep learning model, to conduct detailed inspections of key areas and rapidly survey large areas. This method can be widely used in the maintenance and management of road infrastructure, helping road operators to promptly identify and locate road defects, improve detection efficiency and accuracy, reduce maintenance costs, and prevent further damage, ultimately promoting the sustainable management and development of transportation infrastructure. Summary of the Invention

[0007] In view of this, the present invention provides an automatic detection method for hidden cracks based on multi-feature fusion of radar signals.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] The method for automatically detecting hidden cracks based on multi-feature fusion of radar signals includes the following steps:

[0010] Step 1: Collect the actual highway section and obtain the original ground penetrating radar data

[0011] The acquisition device is connected to the laptop via Ethernet and uses the distance trigger measurement mode to collect the acquisition object;

[0012] Step 2: Processing of GPR signal data and construction of multi-feature GPR dataset

[0013] First, the raw radar data is preprocessed. Then, the Hilbert-Huang transform is used to extract the instantaneous amplitude, instantaneous phase, and instantaneous frequency characteristic signals of the ground penetrating radar signal. Finally, based on the two-dimensional discrete Fourier transform method, the extracted instantaneous amplitude map, instantaneous phase map, and instantaneous frequency map are fused to construct a multi-feature fusion dataset.

[0014] Step 3: Design of a two-stage through-crack detection model based on Swin Transformer optimization

[0015] (3a) Swin Transformer feature extraction network

[0016] Introducing SwinTransformer as the backbone network;

[0017] (3b) Feature Pyramid Network FPN

[0018] The Feature Pyramid Network (FPN) is introduced after the Swin Transformer. It generates feature maps of different depths and scales through a top-down feature extraction network. It also uses upsampling to dynamically adjust the resolution of the deepest feature map and fuse the feature maps of two adjacent layers with different scales.

[0019] (3c) Region Proposal Network (RPN)

[0020] The RPN network generates candidate regions by sliding a window on the feature map, and classifies and regresses these regions to select regions that may contain target objects.

[0021] (3d)ROI pooling optimization

[0022] The ROIAlign algorithm is introduced to divide the candidate region into cells of the same size, and the values ​​of the four center points in each cell are calculated by bilinear interpolation. Then, the ROI output of the same size is obtained by Maxpooling.

[0023] (3e) Smooth L1 loss function

[0024] The Smooth L1 loss function combines the L1 loss function and the L2 loss function;

[0025] Step 4: Design of a one-stage through-crack detection model based on Swin Transformer-YOLOv8 optimization

[0026] First, the Swin Transformer is integrated into the backbone network of YOLOv8, and then the Neck network of YOLOv8 is optimized and the weighted bidirectional feature pyramid (BiFPN) structure is introduced.

[0027] Preferably, in step 1, the acquisition device is connected to the laptop via Ethernet, and the specific configuration is shown in Table 1:

[0028] Table 1 Ground penetrating radar parameter configuration

[0029]

[0030] Preferably, in step 1, the collection objects include municipal roads near the school and multiple roads such as Hedi Road and Shangyuan Road.

[0031] Preferably, in step 2, the preprocessing includes: time-varying zero bias correction, FIR digital filtering, gain signal adjustment and direct wave filtering;

[0032] The time-varying bias correction adopts formula (1):

[0033]

[0034] Among them, x ij Table i column j row data, y ij Represents the data after time-varying bias correction;

[0035] The FIR filter performs digital filtering on the signal using formula (2):

[0036]

[0037] Among them, x n is the input signal, h k is the FIR filter coefficient, y n is the filtered signal;

[0038] The gain adjustment is to adjust the gain of the receiver according to the signal strength through automatic gain setting; the direct wave removal is to eliminate the direct wave that does not contain key information through two-dimensional directed wavelet transformation.

[0039] Preferably, in step 2, the Hilbert-Huang transform extracts the instantaneous amplitude (IA), instantaneous phase (IP) and instantaneous frequency (IF) characteristic signals of the ground penetrating radar signal as shown in formula (3):

[0040]

[0041] Where x(t) is the original signal, is the result of the Hilbert transform, and j is the imaginary unit.

[0042] Preferably, in step 2, the technical process of feature fusion includes weighted superposition of each feature map in the frequency domain, and then performing inverse transformation to obtain a fused feature map.

[0043] Preferably, in step 3, the Swin Transformer Block consists of a multi-layer perception mechanism, a window multi-head self-attention layer, a sliding window multi-head self-attention layer and a normalization layer.

[0044] Preferably, in step 3, the formula of the Smooth L1 loss function is as shown in formula (9):

[0045]

[0046] Here, x represents the difference between the predicted value and the true value. This loss function is more robust when dealing with bounding box regression, which helps improve the stability of the training process and the accuracy of the detection results.

[0047] Preferably, in step 4, BiFPN is fused by skip connection and weighted fusion, and the feature fusion of BiFPN is expressed as shown in formula (10):

[0048] Fusion p ,F c )=α·F p +β·F c Formula (10)

[0049] Among them, F p and F c They represent feature maps from different levels, and α and β are the learned weights.

[0050] Compared with the prior art, the present invention has achieved the following technical effects:

[0051] The present invention significantly improves the accuracy and efficiency of road crack penetration detection by integrating ground-penetrating radar technology and deep learning algorithms. Compared with traditional manual detection methods, the present invention can identify and locate road cracks more quickly and accurately, reducing maintenance time and costs, while also reducing safety risks such as road collapse caused by cracks. In addition, the present invention enhances the model's ability to identify different types of cracks through feature fusion and data augmentation technology, thereby improving the generalization of the detection model. The non-destructive testing method used will not cause damage to the road, which is beneficial to maintaining the integrity of the road and extending its service life. The present invention also provides an intelligent detection method that can continuously monitor road conditions and provide crack detection data in real time, providing a scientific basis for road maintenance decisions and improving road safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A ground penetrating radar crack damage map of the hidden crack automatic detection method based on radar signal multi-feature fusion of the present invention;

[0053] Figure 2 This is the technical roadmap of the method for automatic detection of hidden cracks based on multi-feature fusion of radar signals of the present invention;

[0054] Figure 3This is a graph showing the loss function of a single feature data set for the automatic detection method of hidden cracks based on multi-feature fusion of radar signals according to the present invention;

[0055] Figure 4 This is a feature fusion data set loss function diagram of the hidden crack automatic detection method based on radar signal multi-feature fusion of the present invention;

[0056] Figure 5 These are some test results of the method for automatically detecting hidden cracks based on multi-feature fusion of radar signals according to the present invention;

[0057] (a) Origin; (b) IA; (c) IP; (d) IF; (e) IA+IP; (f) IA+IF; (g) IP+IF; (h) IA+IP+IF. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] The present invention discloses a method for automatically detecting hidden cracks based on multi-feature fusion of radar signals, comprising the following steps:

[0060] Step 1: Collect the actual highway section and obtain the original ground penetrating radar data

[0061] The acquisition equipment is connected to the laptop via Ethernet. The specific configuration is shown in Table 1. The distance trigger measurement mode is used to collect data from the municipal roads near the school, Hedi Road, and Shangyuan Road. By preprocessing the collected ground penetrating radar data, a comparison image of through cracks and non-through cracks on the ground penetrating radar image is finally obtained, as shown in Figure 1. Figure 1 As shown in Figure 3, through these acquisition steps, a dataset containing rich GPR raw data was successfully constructed;

[0062] Table 1 Ground penetrating radar parameter configuration

[0063]

[0064] Step 2: Processing of GPR signal data and construction of multi-feature GPR dataset

[0065] First, the raw radar data is preprocessed, including time-varying zero bias correction, FIR digital filtering, gain signal adjustment, and direct wave filtering.

[0066] The time-varying bias correction adopts formula (1):

[0067]

[0068] Among them, x ij Table i column j row data, y ij Represents the data after time-varying bias correction;

[0069] The FIR filter digitally filters the signal using formula (2):

[0070]

[0071] Among them, x n is the input signal, h k is the FIR filter coefficient, y n is the filtered signal;

[0072] Gain adjustment is achieved by automatic gain setting to adjust the receiver gain according to the signal strength; direct wave removal is achieved by two-dimensional directed wavelet transform to eliminate direct waves that do not contain key information;

[0073] Then, the Hilbert-Huang transform is used to extract the instantaneous amplitude, instantaneous phase and instantaneous frequency characteristic signals of the ground penetrating radar signal. Finally, based on the two-dimensional discrete Fourier transform method, the extracted instantaneous amplitude map, instantaneous phase map and instantaneous frequency map are fused to construct a multi-feature fusion dataset.

[0074] The Hilbert-Huang transform is used to extract the instantaneous amplitude (IA), instantaneous phase (IP) and instantaneous frequency (IF) characteristic signals of the ground penetrating radar signal as shown in formula (3):

[0075]

[0076] Where x(t) is the original signal, is the result of Hilbert transform, j is the imaginary unit;

[0077] The technical process of feature fusion includes weighted superposition of each feature map in the frequency domain, and then inverse transformation to obtain a fused feature map;

[0078] Step 3: Design of a two-stage through-crack detection model based on Swin Transformer optimization

[0079] (3a) Swin Transformer feature extraction network

[0080] Swin Transformer is introduced as the backbone network; its self-attention calculation is limited to the local window, making the computational complexity linear with the image size rather than quadratic; the self-attention mechanism of Swin Transformer is expressed as shown in formula (4):

[0081]

[0082] Among them, Q, K, and V represent query, key, and value respectively. k

[0083] is the dimension of the key. This formula implements self-attention calculation and captures the dependencies between image features.

[0084] Swin TransformerBlock consists of a multi-layer perception mechanism, a window multi-head self-attention layer, a sliding window multi-head self-attention layer, and a normalization layer;

[0085] The output of each part is expressed as formula (5) to (8):

[0086]

[0087]

[0088] (3b) Feature Pyramid Network FPN

[0089] The feature pyramid network (FPN) is introduced after SwinTransformer. It generates feature maps of different depths and scales through a top-down feature extraction network. It also uses upsampling to dynamically adjust the resolution of the deepest feature map and fuse the feature maps of two adjacent layers with different scales.

[0090] (3c) Region Proposal Network (RPN)

[0091] The RPN network generates candidate regions by sliding a window on the feature map, and classifies and regresses these regions to select regions that may contain target objects.

[0092] (3d)ROI pooling optimization

[0093] The ROIAlign algorithm is introduced to divide the candidate region into cells of the same size, and the values ​​of the four center points in each cell are calculated by bilinear interpolation. Then, the ROI output of the same size is obtained by Maxpooling.

[0094] (3e) Smooth L1 loss function

[0095] The Smooth L1 loss function combines the L1 loss function and the L2 loss function;

[0096] The formula of Smooth L1 loss function is shown in formula (9):

[0097]

[0098] Here, x represents the difference between the predicted value and the true value. This loss function is more robust when dealing with bounding box regression, which helps improve the stability of the training process and the accuracy of the detection results.

[0099] Step 4: Design of a one-stage through-crack detection model based on Swin Transformer-YOLOv8 optimization

[0100] First, the Swin Transformer is integrated into the YOLOv8 backbone network. Then, the Neck network of YOLOv8 is optimized and the weighted bidirectional feature pyramid (BiFPN) structure is introduced.

[0101] BiFPN uses skip connections and weighted fusion. The feature fusion representation of BiFPN is shown in formula (10):

[0102] Fusion p ,F c )=α·F p +β·F c Formula (10)

[0103] Among them, F p and F c Represent feature maps from different levels, and α and β are the weights obtained through learning. The technical route of the road surface crack penetration degree detection method of the present invention is as follows: Figure 2 shown.

[0104] Example 1:

[0105] 1. Experimental Conditions

[0106] The experimental hardware environment of this invention is: Intel(R) Xeon(R) Gold 5220R CPU@2.20GHz, NvidiaGeForce RTX 3090; the software environment is: Windows 10 64-bit Professional Edition, cuda_11.3, Cudnn-8.0-windows, Pytorch 1.11.0, Anaconda3-2022.10-Windows-x86_64, Python 3.8.16.

[0107] This experiment used the SGD optimizer as the base iterator, with a weight_decay factor of 0.0001 and a momentum factor of 0.9. Training was performed on a single GPU. The batch size was set to 4 for both the original and improved networks, and the corresponding initial learning rate was set to 0.001. During training, the model first loaded the initial weight file of the Swin Transformer trained on the public ImageNet dataset, and the total number of training iterations was set to 100.

[0108] According to the specific implementation of the present invention, the accuracy (Precision, P), recall (Recall, R), average precision (Average Precision, AP), mean average precision (mean Average Precision, mAP), and PR (Precision-Recall) curve of the model on different data sets are calculated as model evaluation indicators.

[0109] 2. Experimental Setup

[0110] The methods compared in the experiment of the present invention are as follows:

[0111] Faster R-CNN: A classic two-stage detection network, used as a baseline network.

[0112] SSD: One-stage detection network, for comparison.

[0113] YOLOv5: a one-stage detection network using FPN structure, for comparison.

[0114] In the experiment, eight different datasets were constructed, including original signals (Origin), single-feature datasets (IA, IP, IF), and multi-feature fusion datasets (IA+IP, IA+IF, IP+IF, and IA+IP+IF). A two-stage network structure optimized with SwinTransformer was used for training and validation.

[0115] 3. Experimental content

[0116] The experimental results are as follows:

[0117] Single feature datasets: The effects of the original signal and three single feature datasets (IA, IP, IF) on the model performance are compared.

[0118] Multi-feature fusion dataset: The effects of four feature fusion datasets (IA+IP, IA+IF, IP+IF, and IA+IP+IF) on model performance were compared.

[0119] 4. Experimental Results

[0120] Experimental results show that the two-stage network based on Swin Transformer optimization has better performance indicators on the original signal dataset than the single feature dataset, with a mAP accuracy of 0.8477. On the multi-feature fusion dataset, the verification index mAP and R of the IA+IF dataset increased by 3.7% and 2.7% respectively, indicating that the feature fusion dataset can effectively improve the performance of the model. The loss function of the single feature dataset and the feature fusion dataset are as follows: Figure 3 and Figure 4 The specific experimental performance comparison results are shown in Table 2.

[0121] Table 2 Comparison results of different datasets

[0122]

[0123] After the loss function converges during the training process, use the best.pt weight file generated last, load the weight file into the model, and use the test set to test the various indicators of the model. Some test results are as follows Figure 5 shown.

[0124] 5. Results Analysis

[0125] From the experimental results, it can be seen that the two-stage through-crack detection model based on SwinTransformer optimization adopted by the present invention has achieved a higher detection accuracy on the multi-feature fusion dataset, verifying the advancement and effectiveness of the present invention. In particular, on the IA+IF dataset, the model's mAP@0.5 reached 88.48%, which is 3.7% higher than the indicator using the original dataset. In addition, in the ablation experiment of the model structure, the optimized network improved mAP@0.5 by 13% compared with the initial network, and the detection effect was significantly improved. In the comparison between models, the accuracy of the model proposed in this chapter is significantly higher than that of other models, but the model file occupies as much as 518MB of space, and the computational complexity and the number of model parameters are also the largest, which shows that while the present invention improves the detection accuracy, it also increases the complexity of the model.

[0126] Example 2:

[0127] 1. Experimental Conditions

[0128] The software and hardware environments of this experiment are the same as those of Example 1, but the SGD optimizer is used for hyperparameter settings. The initial learning rate is 0.001, weight_decay is 0.0005, momentum is 0.937, the batch size is set to 4, and the total number of training iterations is 100 rounds.

[0129] 2. Experimental Setup

[0130] The experiment compared the performance of YOLOv8, Swin Transformer, and the fusion model Swin Transformer-YOLOv8. The dataset used was the multi-feature dataset of through-cracks constructed in Chapter 3, with a training, validation, and test dataset ratio of 7:2:1. The annotation files were converted from the COCO dataset's JSON format to the txt format required by YOLO.

[0131] 3. Experimental content

[0132] The experiments mainly evaluated the performance of different datasets in the one-stage detection network, especially the IA+IF dataset, and compared the performance of the Swin Transformer-YOLOv8 optimized model with other models.

[0133] 4. Experimental Results

[0134] Tables 3 and 4 detail the performance of each dataset under different evaluation metrics. Table 5 shows the comparative results on the through-crack original signal dataset. Experimental results show that the Swin Transformer-YOLOv8 optimized model achieved a mAP@0.5 score of 86.2% on the IA+IF dataset, 2.4% higher than the original signal dataset. The Swin Transformer-YOLOv8 optimized network significantly alleviated device storage pressure and reduced computational complexity.

[0135] Table 3 Experimental results of mAP@0.5(%) on single feature dataset

[0136]

[0137] Table 4 Experimental results of feature fusion dataset mAP@0.5(%)

[0138]

[0139] Table 5 Comparison of model detection results

[0140]

[0141]

[0142] 5. Results Analysis

[0143] Comprehensive experimental results show that the Swin Transformer-YOLOv8 optimized model significantly improves detection speed while maintaining high detection accuracy, reaching 35.71 frames per second, meeting the requirements of real-time detection. Compared with the two-stage detection model, although accuracy is slightly reduced, detection speed is significantly improved, making it suitable for scenarios requiring rapid detection.

[0144] The present invention aims to solve the problems of low accuracy, poor efficiency and insufficient global information capture in existing methods for detecting the degree of penetration of road cracks. By integrating the instantaneous attribute feature fusion technology of ground penetrating radar images and the deep learning model based on Transformer optimization, the present invention improves the accuracy and efficiency of road crack penetration detection. Compared with the traditional target detection model based on convolutional neural networks, the Transformer model adopted by the present invention can better capture the overall structure and contextual information of the cracks, and enhance the global information modeling capability of crack features. The present invention also introduces the Hilbert-Huang transform to enhance the recognition of crack features, improves the image quality through feature fusion technology, and thus reduces the impact of signal attenuation on the detection results, aiming to provide a more efficient and accurate method for detecting the degree of penetration of road cracks, so as to improve the modernization level of road maintenance and ensure road traffic safety.

[0145] The above description is merely a preferred embodiment of the present invention and does not limit the technical scope of the present invention. Therefore, any minor modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. An automatic detection method for hidden cracks based on multi-feature fusion of radar signals, characterized in that: The steps include: Step 1: Collect the actual highway section and obtain the original ground penetrating radar data The acquisition device is connected to the laptop computer via Ethernet and uses the distance trigger measurement mode to acquire the acquisition object; Step 2: Processing of GPR signal data and construction of multi-feature GPR dataset Firstly, the original radar data is preprocessed, and then the instantaneous amplitude, instantaneous phase and instantaneous frequency characteristic signals of the ground penetrating radar signal are extracted by using the Hilbert-Huang transform. Finally, based on the two-dimensional discrete Fourier transform method, the extracted instantaneous amplitude map, instantaneous phase map and instantaneous frequency map are fused to construct a multi-feature fusion data set. Step 3: Design of a two-stage through crack detection model based on Swin Transformer optimization (3a) Swin Transformer feature extraction network Introducing Swin Transformer as the backbone network; (3b) Feature Pyramid Network FPN The feature pyramid network FPN is introduced after the Swin Transformer. The top-down feature extraction network generates feature maps of different depths and scales. The upsampling method is used to dynamically adjust the resolution of the deepest feature map, and the feature maps of different scales of two adjacent layers are fused. (3c) Region Proposal Network (RPN) The RPN network generates candidate regions by sliding a window on the feature map, and classifies and regresses these regions to select regions that may contain target objects. (3d)ROI pooling optimization The ROI Align algorithm is introduced to divide the candidate region into cells of the same size, and the values ​​of the four center points in each cell are calculated by bilinear interpolation, and then the ROI output of the same size is obtained by Max pooling; (3e) Smooth L1 loss function The Smooth L1 loss function combines the L1 loss function and the L2 loss function; Step 4: Design of a one-stage through crack detection model based on Swin Transformer-YOLOv8 optimization Firstly, Swin Transformer is integrated into the backbone network of YOLOv8, and then the Neck network of YOLOv8 is optimized and the weighted bidirectional feature pyramid (BiFPN) structure is introduced.

2. The method for automatic detection of hidden cracks based on multi-feature fusion of radar signals according to claim 1 is characterized in that: In step 1, the acquisition device is connected to the laptop via Ethernet, and the specific configuration is shown in Table 1: Table 1 Ground Penetrating Radar Parameter Configuration 3. The method for automatic detection of hidden cracks based on multi-feature fusion of radar signals according to claim 1 is characterized in that: In step 1, the collection objects include municipal roads near the school and multiple roads such as Hedi Road and Shangyuan Road.

4. The method for automatic detection of hidden cracks based on multi-feature fusion of radar signals according to claim 1 is characterized in that: In step 2, the preprocessing includes: time-varying zero bias correction, FIR digital filtering, gain signal adjustment and direct wave filtering; The time-varying zero bias correction adopts formula (1): Among them, x ij Table ith column and j row data, y ij Represents the data after time-varying bias correction; The FIR filter performs digital filtering on the signal, using formula (2): Among them, x n is the input signal, h k is the FIR filter coefficient, y n is the signal after filtering; The gain adjustment is to adjust the gain of the receiver according to the signal strength through automatic gain setting; the direct wave removal is to eliminate the direct wave that does not contain key information through two-dimensional directed wavelet transformation.

5. The method for automatic detection of hidden cracks based on multi-feature fusion of radar signals according to claim 1 is characterized in that: In step 2, the Hilbert-Huang transform extracts the instantaneous amplitude (IA), instantaneous phase (IP) and instantaneous frequency (IF) characteristic signals of the ground penetrating radar signal as shown in formula (3): Where x(t) is the original signal, is the result of the Hilbert transform, and j is the imaginary unit.

6. The method for automatic detection of hidden cracks based on multi-feature fusion of radar signals according to claim 1 is characterized in that: In step 2, the technical process of feature fusion includes weighted superposition of each feature map in the frequency domain, and then inverse transformation to obtain a fused feature map.

7. The method for automatic detection of hidden cracks based on multi-feature fusion of radar signals according to claim 1 is characterized in that: In step 3, the Swin Transformer Block consists of a multi-layer perception mechanism, a window multi-head self-attention layer, a sliding window multi-head self-attention layer and a normalization layer.

8. The method for automatic detection of hidden cracks based on multi-feature fusion of radar signals according to claim 1 is characterized in that: In step 3, the formula of the Smooth L1 loss function is shown in formula (9): Among them, x represents the difference between the predicted value and the true value. This loss function is more robust when dealing with bounding box regression, which helps to improve the stability of the training process and the accuracy of the detection results.

9. The method for automatic detection of hidden cracks based on multi-feature fusion of radar signals according to claim 1 is characterized in that: In step 4, BiFPN is fused by skip connection and weighted fusion, and the feature fusion of BiFPN is expressed as shown in formula (10): Fusion(F p ,F c )=α·F p +β·F c expression(10) Among them, F p and F c They represent feature maps from different levels, and α and β are the learned weights.

Images