Method for classifying road cavity morphology based on GPR data

Abstract

The application discloses a road cavity morphology classification method based on GPR data, which comprises the following steps: performing road cavity three-dimensional modeling simulation and obtaining three-dimensional GPR data; extracting and jointly representing GPR morphology information; constructing a cavity morphology classification initial model based on a relation network and training to obtain a cavity morphology classification model; and using the cavity morphology classification model to perform road cavity morphology classification on an actual road. The application deeply mines the internal connection and imaging rules between the cavity morphology and the GPR detection data, explores the cavity morphology interpretation mechanism, uses the relation network as the main architecture of the few-sample learning network, and improves the relation network. The method has the advantages of high reliability and good accuracy in realizing the accurate classification of the road cavity morphology under the premise of less data support.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of ground penetrating radar image data processing, and particularly relates to a road cavity morphology classification method based on GPR data. BACKGROUND

[0002] Road collapse is often caused by the gradual expansion of underground cavities, and underground cavities have strong concealment and suddenness; therefore, large-scale, non-destructive underground cavity morphology evaluation and detection have become the focus of urban management.

[0003] However, non-destructive detection of underground cavities has always been a difficulty in detection. At present, the commonly used detection method is the impulse ground penetrating radar (hereinafter referred to as GPR) detection method. In the process of detecting underground cavities, the identification of the morphology of the cavities is crucial. The size and geometry of the cavity morphology reflect the speed of collapse evolution and the severity, and the cavity morphology can accurately perceive the state of the cavity in the three-dimensional space. Therefore, how to accurately and quickly classify and identify the cavity morphology has become the focus of researchers.

[0004] At present, there are three main problems in the research on the classification of road cavity morphology based on GPR data:

[0005] First, the selection and extraction of morphological features: the complexity of the environment, such as the interference of surrounding pipelines and the backflow of accumulated water, may seriously interfere with the description of the cavity morphology features, and it is impossible to comprehensively describe the reflection characteristics with only one or two features; in addition, the pattern features and morphological attributes obtained still need to be inferred and identified by experienced professionals, and the dependence on manpower is strong;

[0006] Second, accurate classification of GPR cavity morphology: due to the concealment of the cavity itself, the change in size, the variety of morphology, the irregularity of the upper surface, the obvious direction and extension, etc., the cavity morphology analysis faces the difficulty of identification and classification;

[0007] Third, the lack of GPR labeled data set: compared with buried steel bars and underground pipelines, the size and diameter of the cavity with collapse threat are generally larger; it is very laborious to make such cavities in the laboratory, and the cavities made are generally uniform in form, and do not have universality and representativeness. SUMMARY

[0008] The purpose of the present application is to provide a road cavity morphology classification method based on GPR data with high reliability, high precision and less demand for data.

[0009] The road cavity morphology classification method based on GPR data provided by the present application comprises the following steps:

[0010] S1. Adopting three-dimensional forward technology, carrying out three-dimensional modeling simulation of road cavity, and obtaining three-dimensional GPR data;

[0011] S2. According to the three-dimensional GPR data obtained in step S1, extracting and jointly representing the GPR morphological information;

[0012] S3. Constructing an initial model of cavity morphological classification based on a relationship network;

[0013] S4. Using the data obtained in step S2, training the initial model of cavity morphological classification constructed in step S3 to obtain a cavity morphological classification model;

[0014] S5. Using the cavity morphological classification model obtained in step S4, carrying out road cavity morphological classification of actual roads.

[0015] The three-dimensional forward technology in step S1, which is used to carry out three-dimensional modeling simulation of road cavity and obtain three-dimensional GPR data, specifically includes the following steps:

[0016] Setting GPR system parameters: The parameters set include spatial resolution, time window, initial transmitting antenna coordinates, initial receiving antenna coordinates, antenna step distance, measurement point number, excitation signal type and excitation signal frequency;

[0017] Constructing a road simulation structure model: The road structure simulation model is composed of air layer, asphalt layer, concrete layer, gravel layer and soil layer from top to bottom; and the dielectric property parameters of each layer are set;

[0018] Designing four types of cavity morphologies in the constructed road simulation model, including spherical cavity, rectangular cavity, cylindrical cavity and irregular semi-spherical cavity;

[0019] Obtaining three-dimensional GPR cavity data of each type of cavity.

[0020] When setting parameters, the parameters are set according to the following rules:

[0021] The spatial resolution is set to 0.01 m; the time window is set to 14 ns; the initial transmitting antenna coordinates are set to (0.45, 1.0, 0.0) m; the initial receiving antenna coordinates are set to (0.35, 1.0, 0.0) m; the antenna step distance is set to (0.01, 0, 0) m; the number of measuring points is set to 100; the excitation signal type is set to Ricker; the excitation signal frequency is set to 800 MHz; the relative dielectric constant of the air layer is set to 1, and the conductivity is set to 0; the relative dielectric constant of the asphalt layer is set to 6, and the conductivity is set to 0.005 S / m; the relative dielectric constant of the concrete layer is set to 9, and the conductivity is set to 0.05 S / m; the relative dielectric constant of the gravel layer is set to 12, and the conductivity is set to 0.1 S / m.

[0022] The three-dimensional GPR data obtained in step S1 is subjected to GPR morphology information extraction and joint representation according to step S2, specifically including the following steps:

[0023] Parallel to the XOZ plane, two-dimensional data is sequentially extracted from the three-dimensional structure S, and a plurality of profile images S1={B1, B2,..., B n} are obtained by imaging. n Bn is the nth profile image, l y is the length of the three-dimensional structure S in the y-axis direction, and Δy is the spatial step length in the y-axis direction; then the obtained profile images S1 are subjected to equal-interval extraction and are displayed in a stacked manner.

[0024] Parallel to the XOY plane, two-dimensional data is sequentially extracted from the three-dimensional structure S, and a plurality of profile images S2={C1, C2,..., C m} are obtained by imaging. m Cm is the mth profile image, l z is the length of the three-dimensional structure S in the z-axis direction, and Δz is the spatial step length in the z-axis direction; then the obtained profile images S2 are subjected to equal-interval extraction and are displayed in a stacked manner.

[0025] For each model, 8 profile images S1 and 12 profile images S2 are extracted, and are rearranged to obtain a new two-dimensional fusion image I as training data; wherein the 8 profile images S1 are B1-B8 in sequence, and the 12 profile images S2 are C1-C 12 .

[0026] The initial model of the cavity morphology classification based on the relationship network is constructed according to step S3, specifically including the following steps:

[0027] The relationship network is improved to serve as an initial model:

[0028] The improved relationship network comprises an embedding module and a relationship module which are sequentially connected in series;

[0029] The embedding module comprises an embedding first convolutional block, an embedding first max-pooling block, an embedding second convolutional block, an embedding second max-pooling block and an embedding third convolutional block which are sequentially connected in series; wherein the embedding first max-pooling block is a 2*2 max-pooling, and the embedding second max-pooling block is a 2*2 max-pooling; the input of the embedding first convolutional block is the two-dimensional fusion image I obtained in step S2; the output of the embedding module is taken as the input of the relationship module after being processed by feature concatenation again;

[0030] The relationship module comprises a relationship first convolutional block, a relationship first max-pooling block, a relationship second convolutional block, a relationship second max-pooling block, a relationship third convolutional block and a relationship Sigmoid function block which are sequentially connected in series; wherein the relationship first max-pooling block is a 2*2 max-pooling, and the relationship second max-pooling block is a 2*2 max-pooling; the size of the relationship third convolutional block is (H*W*C)*N, there are N groups of convolution kernels with a size of (H*W) and a channel number of C, H is the length of the picture output by the relationship second max-pooling block, W is the width of the picture output by the relationship second max-pooling block, C is the channel number of the picture output by the relationship second max-pooling block, and N is the output dimension; the output of the relationship third convolutional block is N groups of 1*1 data which are then sent to the relationship Sigmoid function block for classification.

[0031] The method provided by the application can realize accurate classification of road cavity morphology with high reliability and good accuracy under the premise of less data support. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The figure is a method flowchart of the application.

[0033] Figure 2 The figure is a model schematic diagram of the road simulation structure model of the application.

[0034] Figure 3 The figure is a cavity simulation model schematic diagram under different morphologies of the application.

[0035] Figure 4 The figure is a schematic diagram of the morphology information extraction of the irregular hemispherical cavity of the application.

[0036] Figure 5 The figure is a schematic diagram of the morphology feature fusion result of the cavity of the application.

[0037] Figure 6 This is a schematic diagram of the network structure of the improved relational network of the present invention.

[0038] Figure 7 This is a schematic diagram comparing the cavity morphology classification results of the present invention. Detailed Implementation

[0039] like Figure 1 The diagram shown illustrates the method flow of this invention: This method for classifying road cavity morphology based on GPR data, provided by this invention, includes the following steps:

[0040] S1. Employ 3D forward modeling technology to perform 3D modeling and simulation of road voids, and acquire 3D GPR data; specifically including the following steps:

[0041] Configure GPR system parameters; the parameters to be configured include spatial resolution, time window, initial transmit antenna coordinates, initial receive antenna coordinates, antenna step distance, measurement point number, excitation signal type, and excitation signal frequency;

[0042] Construct a road simulation structure model (such as Figure 2 As shown): The road structure simulation model consists of an air layer, an asphalt layer, a concrete layer, a gravel layer, and a soil layer from top to bottom; and the dielectric properties of each layer are set.

[0043] Four cavity types were designed in the constructed road simulation model (such as...) Figure 3 As shown), including spherical voids ( Figure 3 (a) ), rectangular hole ( Figure 3 (b) cylindrical cavity Figure 3 (c) and irregular hemispherical cavities ( Figure 3 (d));

[0044] Obtain 3D GPR void data for various types of voids;

[0045] When setting parameters, follow these rules:

[0046] The spatial resolution is set to 0.01 m; the time window is set to 14 ns; the initial transmitting antenna coordinates are set to (0.45, 1.0, 0.0) m; the initial receiving antenna coordinates are set to (0.35, 1.0, 0.0) m; the antenna step distance is set to (0.01, 0, 0) m; the number of measurement points is set to 100; the excitation signal type is set to Ricker; the excitation signal frequency is set to 800 MHz; the relative permittivity of the air layer is set to 1, and the conductivity is set to 0; the relative permittivity of the asphalt layer is set to 6, and the conductivity is set to 0.005 S / m; the relative permittivity of the concrete layer is set to 9, and the conductivity is set to 0.05 S / m; the relative permittivity of the gravel layer is set to 12, and the conductivity is set to 0.1 S / m;

[0047] S2. Extracting and jointly representing GPR morphological information from the three-dimensional GPR data obtained in step S1; specifically including the following steps:

[0048] Parallel to the XOZ plane, two-dimensional data are sequentially extracted from the three-dimensional structure S, and a plurality of profile images S1 = {B1, B2,..., Bn} are obtained by imaging. n}Bn is the nth profile image, n l y is the length of the three-dimensional structure S in the y-axis direction, and Δy is the spatial step length in the y-axis direction; then the obtained profile images S1 are extracted at equal intervals, and are displayed in a stacked manner, as shown in Figure 4 (a);

[0049] Parallel to the XOY plane, two-dimensional data are sequentially extracted from the three-dimensional structure S, and a plurality of profile images S2 = {C1, C2,..., Cm} are obtained by imaging. m}Cm is the mth profile image, m l z is the length of the three-dimensional structure S in the z-axis direction, and Δz is the spatial step length in the z-axis direction; then the obtained profile images S2 are extracted at equal intervals, and are displayed in a stacked manner, as shown in Figure 4 (b);

[0050] For each model, 8 profile images S1 and 12 profile images S2 are extracted, and are rearranged to obtain a new two-dimensional fusion image I as training data; wherein the 8 profile images S1 are B1-B8 in sequence, and the 12 profile images S2 are C1-C 12 . Figure 5 The characteristic fusion results of the inflated cavities under four different morphologies are respectively shown: Figure 5 (a) is the characteristic fusion result under the spherical cavity; Figure 5 ​​(b) is the feature fusion result under the rectangular cavity; Figure 5 (c) is the feature fusion result under the cylindrical cavity; Figure 5 (d) is the feature fusion result under the irregular hemispherical cavity;

[0051] S3. Constructing an initial model of cavity morphology classification based on a relational network; specifically comprising the following steps:

[0052] Improving the relational network as an initial model:

[0053] The improved relational network comprises an embedding module and a relationship module connected in sequence;

[0054] The embedding module comprises an embedding first convolution block, an embedding first max-pooling block, an embedding second convolution block, an embedding second max-pooling block, an embedding third convolution block and an embedding fourth convolution block connected in sequence; wherein the embedding first max-pooling block is a 2*2 max-pooling, and the embedding second max-pooling block is a 2*2 max-pooling; the input of the embedding first convolution block is the two-dimensional fusion image I obtained in step S2; the output of the embedding module is used as the input of the relationship module after being cascaded at the feature level;

[0055] The relationship module comprises a relationship first convolution block, a relationship first max-pooling block, a relationship second convolution block, a relationship second max-pooling block, a relationship third convolution block and a relationship Sigmoid function block connected in sequence; wherein the relationship first max-pooling block is a 2*2 max-pooling, and the relationship second max-pooling block is a 2*2 max-pooling; the size of the relationship third convolution block is (HxWxC)XN, there are N groups of convolution kernels with a size of (HxW) and a channel number C, H is the length of the picture output by the relationship second max-pooling block, W is the width of the picture output by the relationship second max-pooling block, C is the channel number of the picture output by the relationship second max-pooling block, N is the output dimension, there are four categories involved in the present application, so the value of N is 4; the output of the relationship third convolution block is N groups (4 groups) of 1x1 data, which are then sent to the relationship Sigmoid function block for classification; the model structure is as shown in Figure 6

[0056] S4. Training the initial model of cavity morphology classification constructed in step S3 using the data obtained in step S2, to obtain a cavity morphology classification model;

[0057] S5. Using the cavity morphology classification model obtained in step S4 to classify the road cavity morphology of an actual road.

[0058] Figure 7 ​A schematic diagram of the hollow shape classification result of the classification model proposed in the application: only 20 support samples are used to achieve an accuracy of 97.328%. The results are obtained based on the following settings: the backbone network uses Conv64F, the benchmark dataset tieredImageNet, in the form of 4-way 5-shot; Table 1 is the classification result of the relationship network embedded with different backbone networks:

[0059] Table 1 Comparative result schematic table under different backbone networks

[0060] Backbone network 4-way 1-shot 4-way 5-shot Conv64F 78.097 88.934 ResNet12 69.467 72.500 ResNet18 69.926 79.865

[0061] According to Table 1, compared with ResNet12 and ResNet18, Conv64 as the backbone architecture of the relationship network can achieve the best performance on 4-way 5-shot and 4-way 1-shot.

Claims

1. A method for classifying road cavity morphology based on GPR data, comprising the following steps: S1. Using 3D forward modeling technology, perform 3D modeling and simulation of road voids, and obtain 3D GPR data; S2. Based on the three-dimensional GPR data obtained in step S1, extract and jointly characterize the GPR morphological information; S3. Construct an initial model for classifying cavity morphology based on relational networks; specifically including the following steps: The relationship network was improved as an initial model: The improved relational network consists of embedded modules and relational modules connected in sequence; The embedding module includes, in sequence, an embedded first convolutional block, an embedded first max-pooling block, an embedded second convolutional block, an embedded second max-pooling block, and an embedded third convolutional block; wherein, The first max pooling block is embedded as a 2x2 max pooling block, and the second max pooling block is embedded as a 2x2 max pooling block; the input to the first convolutional block is the two-dimensional fused image obtained in step S2. The output of the embedding module is then passed through feature concatenation and used as the input of the relation module. The relation module includes a first convolutional block, a first max-pooling block, a second convolutional block, a second max-pooling block, a third convolutional block, and a sigmoid function block, all connected in sequence. The first max-pooling block uses a 2x2 max pooling function, the second max-pooling block uses a 2x2 max pooling function, and the third convolutional block has a size of [missing information]. There are a total of The group convolution kernel size is and the number of channels is convolution kernel, The length of the image output by the second max-pooling block is given. The width of the image output by the second max-pooling block. This represents the number of channels in the image output by the second max-pooling block. It is the output dimension; the output of the third convolutional block is N sets. The data of size is then fed into the relational Sigmoid function block for classification; S4. Using the data obtained in step S2, train the initial cavity morphology classification model constructed in step S3 to obtain the cavity morphology classification model. S5. Using the cavity morphology classification model obtained in step S4, classify the cavity morphology of actual roads.

2. The road cavity morphology classification method based on GPR data according to claim 1, characterized in that... Step S1, which involves using 3D forward modeling technology to perform 3D modeling and simulation of road cavities and obtain 3D GPR data, specifically includes the following steps: Set GPR system parameters: The parameters to be set include spatial resolution, time window, initial transmit antenna coordinates, initial receive antenna coordinates, antenna step distance, measurement point number, excitation signal type, and excitation signal frequency; Construct a road simulation structure model: The road structure simulation model consists of an air layer, an asphalt layer, a concrete layer, a gravel layer, and a soil layer from top to bottom; and set the dielectric property parameters for each layer. Four cavity shapes were designed in the constructed road simulation model, including spherical cavity, rectangular cavity, cylindrical cavity and irregular hemispherical cavity; Obtain 3D GPR void data for various types of voids.

3. The road cavity morphology classification method based on GPR data according to claim 2, characterized in that... When setting parameters, follow these rules: Spatial resolution is set to 0.01m; time window is set to 14ns; initial transmit antenna coordinates are set to... The unit is meters; the initial receiving antenna coordinates are set to... The unit is meters (m). Antenna step distance set The unit is meters; the measurement point number is set to 100; the excitation signal type is set to Ricker; the excitation signal frequency is set to 800MHz; the relative permittivity of the air layer is set to 1, and the conductivity is set to 0; the relative permittivity of the asphalt layer is set to 6, and the conductivity is set to 0.005S / m; the relative permittivity of the concrete layer is set to 9, and the conductivity is set to 0.05S / m; the relative permittivity of the gravel layer is set to 12, and the conductivity is set to 0.1S / m.

4. The road cavity morphology classification method based on GPR data according to claim 3, characterized in that... Step S2, which involves extracting and jointly representing the morphological information of the three-dimensional GPR data obtained in step S1, specifically includes the following steps: Parallel to the XOZ plane, two-dimensional data are extracted sequentially from the three-dimensional structure S, and several cross-sectional images are obtained. , This is the nth cross-sectional view. , Let S be the length of the three-dimensional structure S along the y-axis. The spatial step size is set along the y-axis; then the cross-sectional view is obtained. Extract at equal intervals and display them in a stacked manner; Parallel to the XOY plane, two-dimensional data are extracted sequentially from the three-dimensional structure S, and several cross-sectional images are obtained. , This is the m-th cross-sectional view. , Let S be the length of the three-dimensional structure S along the z-axis. The spatial step size is set along the z-axis; then the cross-sectional view is obtained. Extract at equal intervals and display them in a stacked manner; For each model, extract 8 cross-sectional views. and 12 cross-sectional drawings The images are then rearranged to form a new two-dimensional fused image. for As training data; including 8 cross-sectional images. In order 12 cross-sectional drawings In order .

Images