Multi-frequency ground penetrating radar data fusion method based on deep learning

Abstract

The invention provides a multi-frequency ground penetrating radar data fusion method based on deep learning. The multi-frequency ground penetrating radar data fusion method comprises the following steps: acquiring multi-frequency ground penetrating radar data by using different center frequency ground penetrating radar devices under the same survey line; performing data preprocessing operation on the acquired multi-frequency ground penetrating radar data; constructing a deep learning model for fusing the multi-frequency ground penetrating radar data, and inputting the multi-frequency data fusion into the fusion model to realize multi-frequency radar data fusion; designing an unsupervised loss function, and training a deep learning model based on the data; and integrating the trained deep learning fusion model into a ground penetrating radar data acquisition system to realize real-time data fusion and processing. Compared with the prior art, the method can integrate the complementary characteristics of the original signals of the multi-frequency radar, and overcomes the limitation and difference of single-frequency data.

Description

Technical Field

[0001] The present invention relates to the technical field of ground penetrating radar signal processing, and in particular to a multi-frequency ground penetrating radar data fusion method based on deep learning. Background Art

[0002] Ground-penetrating radar (GPR) is an important nondestructive testing technology widely used in road inspection, underground pipeline detection, geological exploration, and archaeology. GPRs with different center frequencies have their own advantages: high-frequency GPRs offer higher resolution and can precisely identify shallow structures, while low-frequency GPRs have greater penetration and are suitable for detecting deeper targets. Therefore, single-frequency GPRs struggle to achieve both high resolution and deep penetration, limiting their effectiveness in complex underground environments.

[0003] Fusion of multi-frequency GPR data can compensate for the shortcomings of single-frequency data. However, GPR data from different frequencies exhibit significant differences in spatial resolution, signal-to-noise ratio, and reflectance characteristics. Directly fusing these data can result in information redundancy or loss. Traditional multi-frequency data fusion methods rely primarily on rule-based definitions or linear transformations, lacking the ability to deeply mine complex data features and making it difficult to fully exploit the complementary information of multi-frequency data. In recent years, deep learning technology has demonstrated powerful nonlinear feature extraction capabilities in image processing and signal analysis, providing new insights into complex data fusion. However, in the field of GPR data fusion, effective modeling and fusion of multi-frequency data using deep learning techniques still faces many challenges, such as handling data variability, unsupervised model training, and the need for real-time processing. Therefore, a multi-frequency GPR data fusion method based on deep learning is proposed to improve the quality of data fusion and is of great significance for enhancing the application of GPR in complex environments. Summary of the Invention

[0004] The object of the present invention is to provide a multi-frequency ground penetrating radar data fusion method capable of improving data fusion quality.

[0005] To achieve the above objectives, the present invention proposes a multi-frequency ground penetrating radar data fusion method based on deep learning, comprising the following steps:

[0006] S1: Acquire multi-frequency GPR data using GPR equipment with different center frequencies under the same survey line;

[0007] S2: Perform data preprocessing operations on the acquired multi-frequency ground penetrating radar data;

[0008] S3: constructing a deep learning fusion model for fusing multi-frequency ground penetrating radar data, and fusing the pre-processed multi-frequency ground penetrating radar data into the deep learning fusion model;

[0009] S4: Design an unsupervised loss function and train the deep learning fusion model; the loss function includes the structural similarity loss L ssim , maximum gradient loss L gradient , maximum strength loss L intensity and smoothing loss L smooth One or more weighted combinations of

[0010] S5: Integrate the trained deep learning fusion model into the ground penetrating radar data acquisition system to achieve real-time data fusion and processing.

[0011] Furthermore, in step S1, a ground penetrating radar device with any two or more center frequencies among 100-300 MHz low-frequency ground penetrating radar, 400-600 MHz medium-frequency ground penetrating radar and 800-12000 MHz high-frequency ground penetrating radar is used to scan the same area to obtain multi-frequency ground penetrating radar data.

[0012] Furthermore, in step S2, the preprocessing operation includes one or more of zero point correction, deconvolution, DC offset removal filtering operation, fast low-pass filtering operation and time gain processing.

[0013] Furthermore, in step S3, the deep learning fusion model includes one or more of the following structures: a recurrent neural network model structure (RNN), a convolutional neural network model structure (CNN), and a Transformer model structure.

[0014] Furthermore, in step S3, the method for constructing a deep learning fusion model based on a CNN convolutional neural network is as follows: an encoder-decoder structure is adopted, wherein the encoder is composed of three consecutive convolution modules, and the convolution module includes a 3×3 two-dimensional convolution layer, a LeakyReLU activation function with a negative slope of 0.2, and a batch normalization layer. In the encoder, the feature channels are gradually increased from 16 to 64; the decoder is composed of three consecutive convolution modules, and the feature dimension is gradually reduced from 128 channels to 16 channels; finally, the output layer applies a 1×1 convolution to the final decoder feature to generate a fused data result.

[0015] Furthermore, in step S3, the type of multi-frequency ground penetrating radar data input into the deep learning fusion model includes original data, mapping maps or algorithm-enhanced data.

[0016] Furthermore, in step S3, the method of fusing the preprocessed multi-frequency ground penetrating radar data into the deep learning fusion model includes any one of the following methods: 1) stacking the preprocessed multi-frequency ground penetrating radar signal data and inputting it into the input layer of the model; 2) splicing the preprocessed multi-frequency ground penetrating radar signal data and inputting it into the input layer of the model; 3) adding the preprocessed multi-frequency ground penetrating radar signal data and inputting it into the input layer of the model.

[0017] Furthermore, in step S4, the structural similarity loss L ssim The function is based on structural similarity loss and is used to evaluate the structural similarity between the fused data and the original data, thereby improving the accuracy of the reconstructed data. The formula is:

[0018]

[0019]

[0020] Where Y f Represents fused data, Y1,Y2,…,Y n represents the multi-frequency GPR data to be fused, x and y are the two images to be compared; μ x and μ y is the mean intensity of x and data; and are the variances of x and y, σ xy is the covariance of x and y; C1 and C2 are constants to avoid division by 0;

[0021] Maximum gradient loss L gradient It is used to enhance the ability of fusion data to retain the maximum edge information in the original data. The formula is:

[0022]

[0023] Where, Represents a gradient operation.

[0024] Maximum strength loss L intensity It is used to ensure the consistency between the strongest signal intensity in the fused data and the original data. The formula is:

[0025] Smoothing loss L smooth It is used to constrain the smoothness of the fused data and suppress the influence of high-frequency noise. The formula is:

[0026]

[0027] Where, Represents the gradient value of the fused data at position p; Ω is the total data.

[0028] Furthermore, in step S5, the trained deep learning fusion model is integrated into the ground penetrating radar data acquisition system, and the multi-frequency radar signal data to be processed is obtained through the ground penetrating radar data acquisition system. The fusion method is used to fuse multiple multi-frequency radar signal data into a composite radar signal data, and underground object classification, such as target detection and semantic segmentation, is performed based on the composite radar signal data.

[0029] Furthermore, the fusion method is applied to a multi-frequency ground penetrating radar data fusion device, which includes a memory, a processor, and a program stored in the memory. When the processor executes the program, a multi-frequency ground penetrating radar data fusion method based on deep learning is implemented.

[0030] Furthermore, the fusion method is applied to a storage medium, which stores a program. When the program is executed, the fusion method is implemented.

[0031] Compared with the prior art, the advantages of the present invention are:

[0032] 1. The present invention fully mines and extracts the complementary features of different frequencies through a deep learning model, fuses data to integrate the detailed features of the original signal, and realizes the efficient fusion of multi-frequency ground penetrating radar data, overcoming the limitations and differences of single-frequency data; the fused data simultaneously takes into account the high resolution of shallow targets and the penetration ability of deep targets, which is conducive to the subsequent identification of buried targets.

[0033] 2. The present invention designs a loss function based on unsupervised learning theory, which can effectively process ground penetrating radar data that lacks annotations, providing greater flexibility and adaptability for model training.

[0034] 3. The method of the present invention can effectively improve the quality of data fusion, which is of great significance for improving the application effect of ground penetrating radar in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of the process of a multi-frequency ground penetrating radar data fusion method based on deep learning according to an embodiment of the present invention;

[0036] Figure 2 Schematic diagram of the fusion model framework based on deep learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be further described below.

[0038] This embodiment proposes a multi-frequency ground penetrating radar data fusion method based on deep learning, such as Figure 1 As shown, the method includes the following steps:

[0039] S1: Acquire multi-frequency GPR data using GPR equipment with different center frequencies under the same survey line;

[0040] In this example, two types of GPR equipment were used: a low-frequency GPR with a center frequency of 400 MHz and a high-frequency GPR with a center frequency of 800 MHz. On the same survey line, the two GPRs were used to scan the same area, collecting two sets of multi-frequency GPR data.

[0041] S2: Perform data preprocessing operations on the acquired multi-frequency ground penetrating radar data;

[0042] Specifically, the collected multi-frequency ground-penetrating radar data is zero-corrected to eliminate the effects of initial time delay. A deconvolution algorithm is applied to enhance the target signal and reduce noise interference. Time gain is applied based on the AGC automatic gain algorithm to compensate for the attenuation of deep-layer signals.

[0043] S3: constructing a deep learning fusion model for fusing multi-frequency ground penetrating radar data, and fusing the pre-processed multi-frequency ground penetrating radar data into the deep learning fusion model;

[0044] In this embodiment, a method for constructing a deep learning fusion model based on CNN convolutional neural network is as follows: Figure 2 As shown in the figure, an encoder-decoder structure is adopted. The encoder consists of three consecutive convolution modules, which include a 3×3 two-dimensional convolution layer, a LeakyReLU activation function with a negative slope of 0.2, and a batch normalization layer. In the encoder, the feature channels gradually increase from 16 to 64; the decoder consists of three consecutive convolution modules, which gradually reduce the feature dimension from 128 channels to 16 channels; finally, the output layer applies 1×1 convolution to the final decoder feature to produce a fused data result.

[0045] The preprocessed multi-frequency ground penetrating radar data is fused and input into the above-mentioned deep learning fusion model, wherein the multi-frequency ground penetrating radar data type input into the deep learning fusion model includes original data, mapping atlas and algorithm-enhanced data.

[0046] In this embodiment, the pre-processed multi-frequency ground penetrating radar data is stacked and then input into the above-mentioned deep learning fusion model.

[0047] As another optional implementation, the pre-processed multi-frequency ground penetrating radar signal data may also be input into the input layer of the model through a splicing operation or an addition operation.

[0048] S4: Design unsupervised loss function and train the deep learning fusion model; the loss function includes structural similarity loss L ssim , maximum gradient loss L gradient , maximum strength loss L intensity and smoothing loss L smooth One or more weighted combinations of

[0049] Among them, the structural similarity loss L ssim The function is based on structural similarity loss and is used to evaluate the structural similarity between the fused data and the original data, thereby improving the accuracy of the reconstructed data. The formula is:

[0050]

[0051]

[0052] Where Y f Represents fused data, Y1,Y2,…,Y n represents the multi-frequency GPR data to be fused, x and y are the two images to be compared; μ x and μ y is the mean intensity of x and data; and are the variances of x and y, σ xy is the covariance of x and y; C1 and C2 are constants to avoid division by 0;

[0053] Maximum gradient loss L gradient It is used to enhance the ability of fusion data to retain the maximum edge information in the original data. The formula is:

[0054]

[0055] Where, Represents a gradient operation.

[0056] Maximum strength loss L intensity It is used to ensure the consistency between the strongest signal intensity in the fused data and the original data. The formula is:

[0057] Smoothing loss L smooth It is used to constrain the smoothness of the fused data and suppress the influence of high-frequency noise. The formula is:

[0058]

[0059] Where, Represents the gradient value of the fused data at position p; Ω is the total data.

[0060] S5: Integrate the trained deep learning fusion model into the ground penetrating radar data acquisition system to achieve real-time data fusion and processing.

[0061] The trained deep learning fusion model is integrated into the ground penetrating radar data acquisition system. The multi-frequency radar signal data to be processed is obtained through the ground penetrating radar data acquisition system. The fusion method of this embodiment is used to fuse multiple multi-frequency radar signal data into a composite radar signal data. Based on the composite radar signal data, underground object classification, such as target detection and semantic segmentation, is performed. As an optional implementation method, the target detection model YOLOv7 is used to detect the multi-frequency fusion data.

[0062] The multi-frequency ground penetrating radar data fusion method based on deep learning proposed in this embodiment is applied to an electronic device, which includes one or more processors, a memory, and one or more programs stored in the memory, wherein the one or more programs include instructions for executing the above-mentioned multi-frequency ground penetrating radar data fusion method based on deep learning.

[0063] The above description is merely a preferred embodiment of the present invention and does not limit the present invention in any way. Any person skilled in the art who, without departing from the scope of the present invention, makes any equivalent substitution, modification, or other changes to the technical solution and technical content disclosed in the present invention shall be deemed to be within the scope of the present invention and still fall within the scope of protection of the present invention.

Claims

1. A multi-frequency ground penetrating radar data fusion method based on deep learning, characterized in that: The steps include: S1: Acquire multi-frequency GPR data using GPR equipment with different center frequencies under the same survey line; S2: Perform data preprocessing operations on the acquired multi-frequency ground penetrating radar data; S3: constructing a deep learning fusion model for fusing multi-frequency ground penetrating radar data, and fusing the pre-processed multi-frequency ground penetrating radar data into the deep learning fusion model; S4: Design an unsupervised loss function and train the deep learning fusion model; the loss function includes the structural similarity loss L ssim , maximum gradient loss L gradient , maximum strength loss L intensity and smoothing loss L smooth One or more weighted combinations of S5: Integrate the trained deep learning fusion model into the ground penetrating radar data acquisition system to achieve real-time data fusion and processing.

2. The multi-frequency ground penetrating radar data fusion method based on deep learning according to claim 1, characterized in that: In step S1, a ground penetrating radar device with any two or more center frequencies among 100-300 MHz low-frequency ground penetrating radar, 400-600 MHz medium-frequency ground penetrating radar and 800-12000 MHz high-frequency ground penetrating radar is used to scan the same area to obtain multi-frequency ground penetrating radar data.

3. The multi-frequency ground penetrating radar data fusion method based on deep learning according to claim 1, characterized in that: In step S2, the preprocessing operation includes one or more of zero point correction, deconvolution, DC offset removal filtering operation, fast low-pass filtering operation and time gain processing.

4. The multi-frequency ground penetrating radar data fusion method based on deep learning according to claim 1, characterized in that: In step S3, the deep learning fusion model includes one or more of the following structures: a recurrent neural network model structure RNN, a convolutional neural network model structure CNN, and a Transformer model structure.

5. The multi-frequency ground penetrating radar data fusion method based on deep learning according to claim 1, characterized in that: In step S3, the method for constructing a deep learning fusion model based on a CNN convolutional neural network is as follows: an encoder-decoder structure is adopted, wherein the encoder is composed of three consecutive convolution modules, and the convolution module includes a 3×3 two-dimensional convolution layer, a LeakyReLU activation function with a negative slope of 0.2, and a batch normalization layer. In the encoder, the feature channels are gradually increased from 16 to 64; the decoder is composed of three consecutive convolution modules, and the feature dimension is gradually reduced from 128 channels to 16 channels; finally, the output layer applies a 1×1 convolution to the final decoder feature to generate a fused data result.

6. The multi-frequency ground penetrating radar data fusion method based on deep learning according to claim 1, characterized in that: In step S3, the type of multi-frequency ground penetrating radar data input into the deep learning fusion model includes original data, mapping patterns or algorithm-enhanced data.

7. The multi-frequency ground penetrating radar data fusion method based on deep learning according to claim 1, characterized in that: In step S3, the method of fusing the pre-processed multi-frequency ground penetrating radar data into the deep learning fusion model includes any of the following methods: 1) stacking the pre-processed multi-frequency ground penetrating radar signal data and inputting it into the input layer of the model; 2) The preprocessed multi-frequency ground penetrating radar signal data is spliced and input into the input layer of the model; 3) The preprocessed multi-frequency ground penetrating radar signal data is added and input into the input layer of the model.

8. The multi-frequency ground penetrating radar data fusion method based on deep learning according to claim 1, characterized in that: In step S4, the structural similarity loss L ssim The function is used to evaluate the structural similarity between the fused data and the original data. The formula is: Where Y f Represents fused data, Y1,Y2,…,Y n represents the multi-frequency GPR data to be fused, x and y are the two images to be compared; μ x and μ y is the mean intensity of x and data; and are the variances of x and y, σ xy is the covariance of x and y; C1 and C2 are constants to avoid division by 0; Maximum gradient loss L gradient It is used to enhance the ability of fusion data to retain the maximum edge information in the original data. The formula is: Maximum strength loss L intensity It is used to ensure the consistency between the strongest signal intensity in the fused data and the original data. The formula is: Smoothing loss L smooth It is used to constrain the smoothness of the fused data and suppress the influence of high-frequency noise. The formula is: Where, represents the gradient value of the fused data at position p. Ω represents all the data.

9. The multi-frequency ground penetrating radar data fusion method based on deep learning according to claim 1, characterized in that: In step S5, the trained deep learning fusion model is integrated into the ground penetrating radar data acquisition system, and the multi-frequency radar signal data to be processed is obtained through the ground penetrating radar data acquisition system. The fusion method is used to fuse multiple multi-frequency radar signal data into a composite radar signal data, and underground object classification is performed based on the composite radar signal data.

10. The multi-frequency ground penetrating radar data fusion method based on deep learning according to claim 1, characterized in that: The fusion method is applied to a multi-frequency ground penetrating radar data fusion device, which includes a memory, a processor, and a program stored in the memory. When the processor executes the program, a multi-frequency ground penetrating radar data fusion method based on deep learning is implemented.

Images