A three-dimensional ground penetrating radar deep cavity detection method based on reinforcement learning
By combining reinforcement learning-based methods with 3D ground-penetrating radar data processing, the problem of low efficiency in deep cavity detection was solved, achieving efficient and accurate deep cavity detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH (BEIJING)
- Filing Date
- 2023-06-12
- Publication Date
- 2026-06-26
AI Technical Summary
Existing 3D ground-penetrating radar technology is difficult to detect deep cavities efficiently, and the electromagnetic wave characteristics of deep cavities are weak, resulting in low data processing efficiency.
A reinforcement learning-based approach is adopted. By pre-training deep cavity and normal soil classifiers, combined with 3D step search and region adjustment actions, the detection of deep cavity regions is iteratively optimized. Features are extracted using 3D convolution and transposed convolution, and the probability of cavity is determined using the Softmax formula.
It achieves efficient and accurate detection of deep cavities, improves data processing efficiency and detection accuracy, and adapts to the complex electromagnetic wave characteristics of deep cavities.
Smart Images
Figure CN116774153B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional ground-penetrating radar signal processing technology, and more specifically to a three-dimensional ground-penetrating radar intelligent detection method for deep cavities based on reinforcement learning. Background Technology
[0002] Urban roads are crucial municipal infrastructure for urban development, and road collapse has become a major hazard to urban road infrastructure. Ground-penetrating radar (GPR), as a fast, efficient, and non-destructive detection method, has been widely used in detecting underground defects in urban roads. However, commonly used array GPRs, due to their limited number of radar channels, cannot fully reflect the information of the underground space being detected. 3D GPR, by arranging multiple sets of parallel antennas, can obtain the underground 3D spatial data volume in a single detection. Its detection width, detection accuracy, and information acquisition angle are all superior to traditional array GPRs.
[0003] The significantly increased data volume of 3D ground-penetrating radar compared to 2D ground-penetrating radar places higher demands on data processing efficiency, making manual processing insufficient for practical engineering needs. Furthermore, deep cavities in roads exhibit significantly fewer diffraction and multiple wave characteristics compared to shallow cavities, resulting in greater electromagnetic wave energy attenuation. Currently, there is no specific method for detecting deep cavities using 3D ground-penetrating radar. Summary of the Invention
[0004] In view of this, the present invention provides a method for detecting deep cavities using 3D ground-penetrating radar (GPR) based on reinforcement learning. By using 3D GPR data of deep normal soil and deep cavities, classifiers for deep cavities and normal soil are pre-trained to obtain the probability that input data belongs to either deep cavities or soil. Then, by using a search box with an initialized step size, the method progressively searches the deep 3D GPR data to be detected. The three regions with the highest probability of being deep cavities are selected. Using these three regions as benchmarks, the 3D region adjustment action is iteratively applied, and the decision to retain the current action is based on the change in the probability of the current region belonging to a cavity. This process continues until the probability of the current region belonging to a deep cavity stabilizes, at which point the region with the highest probability is selected as the detection result of the deep cavity and output.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] S1: Pre-training of deep cavity and normal soil classifiers, the specific steps are as follows:
[0007] S11: Adjust the size of the 3D ground-penetrating radar data containing deep cavities to 64*64*64, and label it as "cavities";
[0008] S12: Adjust the size of the 3D ground-penetrating radar data of deep normal soil to 64*64*64, and label it as "normal soil";
[0009] S13: Using the data from S11 and S12, pre-train the deep cavity classifier so that it can correctly output the probability that the input data belongs to deep cavity or normal soil. The feature extraction part of the deep cavity classifier and soil discriminator is completed by 3D convolution and 3D transposed convolution. The probability of belonging to cavity or soil is given by the Softmax formula, which is as follows:
[0010]
[0011] Where z i Let C be the output value of the i-th class, and C be the number of classes. The final output value is a probability distribution between 0 and 1.
[0012] S2: 3D step search and 3D cavity region regression, the specific steps are as follows:
[0013] S21: Using a three-dimensional stepping method, traverse the deep three-dimensional ground-penetrating radar data to be detected. The dimensions of the three-dimensional search window in the x, y, and z directions are x_w, y_w, and z_w, respectively, and the movement step sizes of the window in the x, y, and z directions are x_s, y_s, and z_s, respectively. The dimensions of the deep three-dimensional radar data volume to be searched are x_L, y_L, and z_L. Then, the number of regions to be detected, D_N, generated by the three-dimensional stepping search satisfies:
[0014]
[0015] S22: Input the D_N regions to be detected generated in S21 into the pre-trained deep cavity and normal soil classifier in S1 to obtain the probability of D_N regions belonging to deep cavities, and select the 3 regions with the highest probability of deep cavities.
[0016] S23: Using the three regions in S22 as a reference, iteratively adjust the range of the current data region in the x, y, and z directions. The adjustment step sizes in the x, y, and z directions are x_a, y_a, and z_a, respectively. After adjusting the data range, the data size D_S is:
[0017]
[0018] Re-input the data into the deep cavity and normal soil classifier in S1. If the probability of the current area being a cavity increases, retain the current data range adjustment action; otherwise, cancel the current data range adjustment and continue to adjust the data range in the next direction.
[0019] S24: The iteration stops when the probability that the current region belongs to a deep cavity remains stable. The condition for the probability to remain stable is that after the action adjustment, the probability that the region belongs to a cavity decreases five times consecutively. The condition for stopping the iteration is as follows:
[0020]
[0021] Where p_C is the probability of belonging to a cavity, p_C1 to p_C5 are the probabilities of the region belonging to a cavity in the previous five times. The probabilities of the three regions selected in S23 belonging to cavities are compared, and the region with the highest probability is selected as the detection result of the deep cavity and output. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0023] Figure 1 The attached figure is a flowchart of a three-dimensional ground-penetrating radar method for detecting deep cavities provided by the present invention.
[0024] Figure 2 The attached diagram shows a schematic diagram of the deep cavity and normal soil discriminator.
[0025] Figure 3 The attached figure is a schematic diagram of the three-dimensional region adjustment action designed in this invention.
[0026] Specific implementation method
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] See appendix Figure 1 This invention discloses a method for detecting deep cavities using three-dimensional ground-penetrating radar based on reinforcement learning, comprising:
[0029] S1: Pre-training of deep cavity and normal soil classifiers, the specific steps are as follows:
[0030] S11: Adjust the size of the 3D ground-penetrating radar data containing deep cavities to 64*64*64, and label it as "cavities";
[0031] S12: Adjust the size of the 3D ground-penetrating radar data of deep normal soil to 64*64*64, and label it as "normal soil";
[0032] S13: Using the data from S11 and S12, pre-train the deep cavity classifier so that it can correctly output the probability that the input data belongs to deep cavity or normal soil. The feature extraction part of the deep cavity classifier and soil discriminator is completed by 3D convolution and 3D transposed convolution. The probability of belonging to cavity or soil is given by the Softmax formula, which is as follows:
[0033]
[0034] Where z i Let C be the output value of the i-th class, and C be the number of classes. The final output value is a probability distribution between 0 and 1.
[0035] S2: 3D step search and 3D cavity region regression, the specific steps are as follows:
[0036] S21: Using a three-dimensional stepping method, traverse the deep three-dimensional ground-penetrating radar data to be detected. The dimensions of the three-dimensional search window in the x, y, and z directions are x_w, y_w, and z_w, respectively, and the movement step sizes of the window in the x, y, and z directions are x_s, y_s, and z_s, respectively. The dimensions of the deep three-dimensional radar data volume to be searched are x_L, y_L, and z_L. Then, the number of regions to be detected, D_N, generated by the three-dimensional stepping search satisfies:
[0037]
[0038] S22: Input the D_N regions to be detected generated in S21 into the pre-trained deep cavity and normal soil classifier in S1 to obtain the probability of D_N regions belonging to deep cavities, and select the 3 regions with the highest probability of deep cavities.
[0039] S23: Using the three regions in S22 as a reference, iteratively adjust the range of the current data region in the x, y, and z directions. The adjustment step sizes in the x, y, and z directions are x_a, y_a, and z_a, respectively. After adjusting the data range, the data size D_S is:
[0040]
[0041] Re-input the data into the deep cavity and normal soil classifier in S1. If the probability of the current area being a cavity increases, retain the current data range adjustment action; otherwise, cancel the current data range adjustment and continue to adjust the data range in the next direction.
[0042] S24: The iteration stops when the probability that the current region belongs to a deep cavity remains stable. The condition for the probability to remain stable is that after the action adjustment, the probability that the region belongs to a cavity decreases five times consecutively. The condition for stopping the iteration is as follows:
[0043]
[0044] Where p_C is the probability of belonging to a cavity, p_C1 to p_C5 are the probabilities of the region belonging to a cavity in the previous five times. The probabilities of the three regions selected in S23 belonging to cavities are compared, and the region with the highest probability is selected as the detection result of the deep cavity and output.
[0045] Figure 2 The diagram shows the discriminator structure for deep cavities and normal soil. During discriminator pre-training, the input data consists of 3D radar data of deep cavities and normal soil. The discriminator uses 3D convolutional kernels moving in the x, y, and z directions to extract data features. The discriminator convolutional kernel size is (7, 7, 7), and the stride is (1, 1, 1). Convolution on a large overlapping window ensures information stability and spatial relevance of the convolution results. To avoid information loss due to multiple convolutions, transposed convolutions are used for upsampling in the sixth and seventh layers of the discriminator network before further convolution. The parameters of the transposed convolution are learnable. By adding a transposed convolution structure to the network structure to fuse the features from previous multiple convolutions, the network parameters can be optimized, target features can be accurately extracted, and the difference between cavities and normal soil can be precisely provided. After convolution, the hole value discriminator further compresses the data features through three fully connected layers, and finally outputs the hole similarity and normal soil similarity of the current data through Softmax, which serves as the basis for subsequent operations.
[0046] Figure 3 The diagram shows a schematic of the three-dimensional region adjustment action designed in this invention. As shown, the three-dimensional region adjustment action consists of adjustments in three directions: horizontal, vertical, and forward / backward, adjusting the size of the current region.
[0047] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for detecting deep cavities using three-dimensional ground-penetrating radar based on reinforcement learning, characterized in that, include: S1: Pre-training of deep cavity and normal soil classifiers, the specific steps are as follows: S11: Adjust the size of the 3D ground-penetrating radar data containing deep cavities to 64*64*64, and label it as "cavities"; S12: Adjust the size of the 3D ground-penetrating radar data of deep normal soil to 64*64*64, and label it as "normal soil"; S13: Using the data from S11 and S12, pre-train the deep cavity classifier so that it can correctly output the probability that the input data belongs to deep cavity or normal soil. The feature extraction part of the deep cavity classifier and soil discriminator is completed by 3D convolution and 3D transposed convolution. The probability of belonging to cavity or soil is given by the Softmax formula, which is as follows: Where z i Let C be the output value of the i-th class, and C be the number of classes. The final output value is a probability distribution between 0 and 1. S2: 3D step search and 3D cavity region regression, the specific steps are as follows: S21: Using a three-dimensional stepping method, traverse the deep three-dimensional ground-penetrating radar data to be detected. The dimensions of the three-dimensional search window in the x, y, and z directions are x_w, y_w, and z_w, respectively, and the movement step sizes of the window in the x, y, and z directions are x_s, y_s, and z_s, respectively. The dimensions of the deep three-dimensional radar data volume to be searched are x_L, y_L, and z_L. Then, the number of regions to be detected, D_N, generated by the three-dimensional stepping search satisfies: S22: Input the D_N regions to be detected generated in S21 into the pre-trained deep cavity and normal soil classifier in S1 to obtain the probability of D_N regions belonging to deep cavities, and select the 3 regions with the highest probability of deep cavities. S23: Using the three regions in S22 as a reference, iteratively adjust the range of the current data region in the x, y, and z directions. The adjustment step sizes in the x, y, and z directions are x_a, y_a, and z_a, respectively. After adjusting the data range, the data size D_S is: Re-input the data into the deep cavity and normal soil classifier in S1. If the probability of the current area being a cavity increases, retain the current data range adjustment action; otherwise, cancel the current data range adjustment and continue to adjust the data range in the next direction. S24: The iteration stops when the probability that the current region belongs to a deep cavity remains stable. The condition for the probability to remain stable is that after the action adjustment, the probability that the region belongs to a cavity decreases five times consecutively. The condition for stopping the iteration is as follows: Where p_C is the probability of belonging to a cavity, p_C1 to p_C5 are the probabilities of the region belonging to a cavity in the previous five times. The probabilities of the three regions selected in S23 belonging to cavities are compared, and the region with the highest probability is selected as the detection result of the deep cavity and output.