A ground penetrating radar data sequence modeling encoding method and system

Abstract

The present disclosure belongs to the field of ground penetrating radar technology, and in particular, relates to a ground penetrating radar data sequence modeling coding method and system. The method comprises: obtaining raw radar data, intercepting the raw radar data using a sliding window, and determining a training sample; the training sample includes a plurality of window data; constructing an initial encoder neural network according to the correlation and signal characteristics between the window data in the training sample; and training the initial encoder neural network using the training sample to determine a target encoder neural network. The present disclosure can well model the correlation or statistical relationship between A-scans, and reduce coding noise.

Description

Technical Field

[0001] The present invention belongs to the technical field of ground penetrating radar, and in particular relates to a ground penetrating radar data sequence modeling encoding method and system. Background Art

[0002] Among the existing tunnel detection technologies, ground penetrating radar (GPR) has the characteristics of non-destructive, fast speed and high accuracy. It has been widely used in various fields such as urban infrastructure detection and road and bridge detection to detect unknown objects in the shallow underground.

[0003] The radar transmitter continuously emits electromagnetic wave signals as the radar moves (or over time), and the receiver receives the returned signals. The echo signal at the same position is the A-scan time-series signal, and multiple A-scan time-series signals are stacked to form the B-scan time-space signal of the ground penetrating radar. In tunnel detection, underground targets show hyperbolic and three-phase morphological features in the B-scan time-space signal. Using these characteristics, digital image analysis, machine learning, mathematical models, and deep learning methods can be used to interpret target information such as steel bars, arches, and gaps in radar data.

[0004] In radar data analysis tasks, deep learning methods generally use convolution algorithms to extract features from radar B-scan data in the form of matrix modeling (image recognition). However, convolution operations are local receptive field feature extraction operators that cannot capture long-range features of data and are not fully adapted to radar data analysis tasks. Currently, artificial intelligence methods for capturing long-range features of data are mainly sequence data modeling methods. For example, algorithms such as RNN, LSTM, and Transformer are used to identify targets such as steel bars, arches, and gaps in radar data by constructing sequence labeling tasks.

[0005] As we all know, the encoding method of sequence units is very important for modeling sequence data. Existing sequence encoding methods can be divided into model-driven encoding and data-driven encoding. Based on the model-driven encoding method, the A-scan data can be encoded using wavelet transform according to the time-frequency characteristics of the electromagnetic wave signal. However, the dielectric constant of the underground medium is irregular, and there are various electromagnetic interferences, which cannot be accurately mathematically modeled. At present, artificial intelligence is developing rapidly, and data-driven encoding methods are more promising. Foreign researchers have used one-dimensional (1D-CNN) convolution to encode radar data. However, 1D convolution encodes each A-scan or each group of A-scans, and cannot well model the correlation or statistical relationship between A-scans, and the encoding noise is large.

[0006] It can be seen that an important reason why the accuracy and robustness of the current intelligent interpretation of ground penetrating radar data are limited is the lack of a good data encoding method. Summary of the invention

[0007] In view of the above problems, the present disclosure provides a ground penetrating radar data sequence modeling and encoding method, the method comprising:

[0008] Acquire original radar data, intercept the original radar data using a sliding window, and determine a training sample; the training sample includes a plurality of window data;

[0009] constructing an initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples;

[0010] The initial encoder neural network is trained using the training samples to determine a target encoder neural network.

[0011] Preferably, obtaining original radar data, intercepting the original radar data using a sliding window, and determining training samples include:

[0012] Using a sliding window to intercept the original radar data to determine an initial data block;

[0013] Performing data enhancement on the initial data block to determine an enhanced data block;

[0014] The training sample includes the initial data block and the enhanced data block.

[0015] Preferably, constructing an initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples comprises:

[0016] Determining the network architecture of the initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples;

[0017] The network architecture of the initial encoder neural network includes a window selection module, a feature dimension enhancement module, a self-attention module and an output module.

[0018] Preferably, the window selection module is used to extract each of the window data and add a weight parameter to the window data to determine the participation degree of each of the window data.

[0019] Preferably, the feature dimensionality increasing module is used to increase the dimensional expression of each of the window data, and map the sampling points of the window data to a high-dimensional space vector representation.

[0020] Preferably, the self-attention module is used to introduce a coding structure and perform self-attention learning.

[0021] Preferably, the loss function of the initial encoder neural network is The expression is as follows:

[0022] ;

[0023] in, To adjust the parameters; Input data for the corresponding model The output module mapping vector of Yes and Similarly, after data enhancement, the model input data is mapped to a vector through the output module; Input data for the corresponding model The output module mapping vector.

[0024] Preferably, using the training samples to train the initial encoder neural network to determine a target encoder neural network comprises:

[0025] The initial data block and the enhanced data block are used as inputs of the initial encoder neural network, and the initial encoder neural network is trained with the goal of minimizing the loss function value of the initial encoder neural network to obtain the target encoder neural network.

[0026] The present disclosure also proposes a ground penetrating radar data sequence modeling and coding system, the system comprising:

[0027] An interception module is used to obtain raw radar data, intercept the raw radar data using a sliding window, and determine a training sample; the training sample includes a plurality of window data;

[0028] A construction module, used to construct an initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples;

[0029] A training module is used to train the initial encoder neural network using the training samples to determine a target encoder neural network.

[0030] Preferably, the interception module is used to obtain raw radar data, intercept the raw radar data using a sliding window, and determine training samples, including:

[0031] The interception module is used to intercept the original radar data using a sliding window to determine an initial data block;

[0032] Performing data enhancement on the initial data block to determine an enhanced data block;

[0033] The training sample includes the initial data block and the enhanced data block.

[0034] Preferably, the construction module is used to construct an initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples, including:

[0035] The construction module is used to determine the network architecture of the initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples;

[0036] The network architecture of the initial encoder neural network includes a window selection module, a feature dimension enhancement module, a self-attention module and an output module.

[0037] Preferably, the training module is used to train the initial encoder neural network using the training samples to determine the target encoder neural network, including:

[0038] The training module is used to use the initial data block and the enhanced data block as inputs of the initial encoder neural network, and to train the initial encoder neural network with the goal of minimizing the loss function value of the initial encoder neural network to obtain the target encoder neural network.

[0039] The present disclosure has the following beneficial effects:

[0040] The present invention obtains original radar data, intercepts the original radar data using a sliding window, and determines training samples; constructs an initial encoder neural network according to the correlation and signal characteristics between window data in the training samples; trains the initial encoder neural network using the training samples to determine a target encoder neural network; the correlation or statistical relationship between A-scans can be well modeled, and coding noise is reduced.

[0041] Other features and advantages of the present disclosure will be described in the following description, and partly become apparent from the description, or be understood by implementing the present disclosure. The purpose and other advantages of the present disclosure can be realized and obtained by the structures indicated in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present disclosure or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 A diagram showing a method for modeling and encoding a ground penetrating radar data sequence in an embodiment of the present disclosure is shown;

[0044] Figure 2 The neural network contrast learning structure in the embodiment of the present disclosure is shown;

[0045] Figure 3 A structural diagram of a neural network encoder in an embodiment of the present disclosure is shown;

[0046] Figure 4 A structural diagram of a window selection module in an embodiment of the present disclosure is shown;

[0047] Figure 5 The structure diagram of the dimension-raising module in the embodiment of the present disclosure is shown;

[0048] Figure 6 The structure diagram of the self-attention module in the embodiment of the present disclosure is shown;

[0049] Figure 7 The output module structure diagram in the embodiment of the present disclosure is shown;

[0050] Figure 8 An example of a neural network encoder structure diagram in an embodiment of the present disclosure is shown;

[0051] Fig. 9 A diagram of a ground penetrating radar data sequence modeling coding system in an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0052] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as being limited to the examples set forth herein; on the contrary, these embodiments are provided so that the present disclosure will be more comprehensive and complete, and the concepts of the example embodiments are fully conveyed to those skilled in the art. The described features, structures, or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced while omitting one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known technical solutions are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0053] In addition, the accompanying drawings are only schematic illustrations of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and thus their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or implemented in one or more hardware units or integrated circuits, or implemented in different networks and / or processor devices and / or microcontroller devices.

[0054] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the steps. For example, some steps may be decomposed, while some steps may be combined or partially combined, so the actual execution order may change according to the actual situation.

[0055] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein, for example.

[0056] In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or sub-modules is not necessarily limited to those steps or sub-modules explicitly listed, but may include other steps or sub-modules not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0057] like Figure 1 As shown, the present disclosure proposes a ground penetrating radar data sequence modeling and encoding method, the method comprising:

[0058] Acquire original radar data, intercept the original radar data using a sliding window, and determine a training sample; the training sample includes a plurality of window data;

[0059] constructing an initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples;

[0060] The initial encoder neural network is trained using the training samples to determine a target encoder neural network.

[0061] Specifically, obtaining original radar data, intercepting the original radar data using a sliding window, and determining training samples include:

[0062] Using a sliding window to intercept the original radar data to determine an initial data block;

[0063] Performing data enhancement on the initial data block to determine an enhanced data block;

[0064] The training sample includes the initial data block and the enhanced data block.

[0065] Specifically, according to the correlation and signal characteristics between the window data in the training samples, an initial encoder neural network is constructed, including:

[0066] Determining the network architecture of the initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples;

[0067] The network architecture of the initial encoder neural network includes a window selection module, a feature dimension enhancement module, a self-attention module and an output module.

[0068] Specifically, the window selection module is used to extract each of the window data and add a weight parameter to the window data to determine the participation degree of each of the window data.

[0069] Specifically, the feature dimensionality increasing module is used to increase the dimensional expression of each window data, and map the sampling points of the window data to a high-dimensional space vector representation.

[0070] Specifically, the self-attention module is used to introduce a coding structure and perform self-attention learning.

[0071] Specifically, the loss function of the initial encoder neural network is The expression is as follows:

[0072] ;

[0073] in, To adjust the parameters; Corresponding model input data The output module maps out a vector; Yes and Similarly, after data enhancement, the model input data is mapped to a vector through the output module; Input data for the corresponding model The output module mapping vector.

[0074] Specifically, the initial encoder neural network is trained using the training sample to determine a target encoder neural network, including:

[0075] The initial data block and the enhanced data block are used as inputs of the initial encoder neural network, and the initial encoder neural network is trained with the goal of minimizing the loss function value of the initial encoder neural network to obtain the target encoder neural network.

[0076] In this embodiment, the encoder neural network may also adopt other training methods such as supervised learning, adversarial learning, semi-supervised learning, etc.

[0077] The present invention fully considers the spatiotemporal characteristics of ground penetrating radar data and adopts a sliding window data sampling method to construct a coding network that can model the correlation between A-scan data in the window. The coding network is trained by using a self-supervised learning method, such as Figure 2This technique is a data encoding method for radar data sequence modeling.

[0078] The steps of this program are as follows:

[0079] Step 1: Take the radar B-scan data along the survey line using a sliding window to extract data blocks and perform data enhancement on the data blocks. , data augmentation data block record , which is used as the training sample of the encoding network.

[0080] Step 2: Construct an encoder neural network based on the correlation between A-scans in the window and the A-scan signal characteristics.

[0081] Step 3: and As training samples, the training of the encoding network is completed. This encoding network encodes and maps the radar data.

[0082] The specific steps are as follows:

[0083] In step 1, the radar data B-scan data is intercepted along the survey line using a sliding window of size M to capture the B-scan data block. , where a single data block The shape is (N represents the number of A-scan data sampling points, and M represents the number of A-scan intercepts). Perform data enhancement operations, including up and down translation, random noise and drop operations, to obtain enhanced data .

[0084] In step 2, the encoder neural network is constructed based on the correlation between A-scans in the window and the A-scan signal characteristics. The network architecture includes a window selection module, a feature dimension upgrade module, a self-attention module, and an output module. Figure 3-Figure 7 As shown, Figure 8 It is a specific example diagram.

[0085] The window selection module will extract each A-scan data , increase the w weight parameter, and the neural network adaptively confirms the participation degree of each A-scan data.

[0086] (1)

[0087] in, is the offset value.

[0088] The feature dimension-enhancing module increases the dimension expression of each A-scan and converts the A-scan sampling points Map it to a high-dimensional space vector representation of size N, and finally perform addition and merging operations.

[0089] (2)

[0090] in, A mapping method from 1-dimensional data to n-dimensional data through a fully connected operation; The mapped value.

[0091] The self-attention module introduces the encoding structure in the T5 architecture for self-attention learning.

[0092] Mapping vectors to output modules , the loss function formula of the model is as follows:

[0093] (3)

[0094] in, To adjust the parameters.

[0095] In step 3, from and Take the input of the training sample composition encoder network, where right is a positive sample, For other samples, they are negative samples, and the encoder training is completed using massive data. This encoding network maps the B-scan data from MxN to MxV, M remains unchanged, and the sequence annotation label is not affected. The encoded data and annotation can be directly used in the sequence modeling algorithm.

[0096] Example

[0097] like Figure 8 As shown in the figure, about 150G of tunnel radar data is used for self-supervised training, the sliding window is set to 500, and it is divided into 160782 data blocks of 500*512 size, and the model training is carried out for 87 rounds.

[0098] The trained encoder was validated using a sequence labeling model built based on Transformer. About 10G of radar data containing steel bars was used, which was divided into 12,381 data blocks of 512*512 size. 2,476 of them were extracted as validation sets. The recognition results are shown in Table 1.

[0099] Table 1

[0100]

[0101] like Fig. 9 As shown, the present disclosure also proposes a ground penetrating radar data sequence modeling and coding system, the system comprising:

[0102] An interception module is used to obtain raw radar data, intercept the raw radar data using a sliding window, and determine a training sample; the training sample includes a plurality of window data;

[0103] A construction module, used to construct an initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples;

[0104] A training module is used to train the initial encoder neural network using the training samples to determine a target encoder neural network.

[0105] Those skilled in the art should understand that although the present disclosure is described in detail with reference to the aforementioned embodiments, it is still possible to modify the technical solutions recorded in the aforementioned embodiments, or to make equivalent replacements for some of the technical features therein; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. A ground penetrating radar data sequence modeling and encoding method, characterized in that: The method comprises: Acquire original radar data, intercept the original radar data using a sliding window, and determine a training sample; the training sample includes a plurality of window data; According to the correlation and signal characteristics between the window data in the training samples, an initial encoder neural network is constructed, including: according to the correlation and signal characteristics between the window data in the training samples, the network architecture of the initial encoder neural network is determined; the network architecture of the initial encoder neural network includes a window selection module, a feature dimension upgrade module, a self-attention module and an output module; the window selection module is used to extract each of the window data and add weight parameters to the window data to determine the participation of each of the window data; the feature dimension upgrade module is used to add dimensional expression to each of the window data, and map the sampling points of the window data to a high-dimensional space vector representation; the self-attention module is used to introduce a coding structure and perform self-attention learning; The initial encoder neural network is trained using the training samples to determine a target encoder neural network.

2. The ground penetrating radar data sequence modeling and encoding method according to claim 1, characterized in that: Acquiring original radar data, intercepting the original radar data using a sliding window, and determining training samples, including: Using a sliding window to intercept the original radar data to determine an initial data block; Performing data enhancement on the initial data block to determine an enhanced data block; The training sample includes the initial data block and the enhanced data block.

3. The ground penetrating radar data sequence modeling and encoding method according to claim 2, characterized in that: The loss function of the initial encoder neural network is The expression is as follows: ； in, To adjust the parameters; Input data for the corresponding model The output module mapping vector of Yes and Similarly, after data enhancement, the model input data is mapped to a vector through the output module; Input data for the corresponding model The output module mapping vector.

4. The method for modeling and encoding ground penetrating radar data sequence according to claim 3, characterized in that: Using the training samples to train the initial encoder neural network to determine a target encoder neural network includes: The initial data block and the enhanced data block are used as inputs of the initial encoder neural network, and the initial encoder neural network is trained with the goal of minimizing the loss function value of the initial encoder neural network to obtain the target encoder neural network.

5. A ground penetrating radar data sequence modeling and coding system, characterized in that: The system comprises: An interception module is used to obtain raw radar data, intercept the raw radar data using a sliding window, and determine a training sample; the training sample includes a plurality of window data; A construction module, used to construct an initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples, including: the construction module is used to determine the network architecture of the initial encoder neural network according to the correlation and signal characteristics between the window data in the training samples; the network architecture of the initial encoder neural network includes a window selection module, a feature dimension upgrade module, a self-attention module and an output module; the window selection module is used to extract each of the window data and add weight parameters to the window data to determine the participation of each of the window data; the feature dimension upgrade module is used to add dimensional expression to each of the window data, and map the sampling points of the window data to a high-dimensional space vector representation; the self-attention module is used to introduce a coding structure and perform self-attention learning; A training module is used to train the initial encoder neural network using the training samples to determine a target encoder neural network.

6. The ground penetrating radar data sequence modeling and coding system according to claim 5, characterized in that: The interception module is used to obtain raw radar data, intercept the raw radar data using a sliding window, and determine a training sample, including: The interception module is used to intercept the original radar data using a sliding window to determine an initial data block; Performing data enhancement on the initial data block to determine an enhanced data block; The training sample includes the initial data block and the enhanced data block.

7. The ground penetrating radar data sequence modeling and coding system according to claim 6, characterized in that: The training module is used to train the initial encoder neural network using the training samples to determine the target encoder neural network, including: The training module is used to use the initial data block and the enhanced data block as inputs of the initial encoder neural network, and to train the initial encoder neural network with the goal of minimizing the loss function value of the initial encoder neural network to obtain the target encoder neural network.

Images