A radar intelligent identification method for bridge steel bars

Abstract

The present invention provides a radar intelligent identification method for bridge steel bars, which changes the dielectric constant and conductivity of the bridge structure, and obtains its response characteristics through forward modeling; collects typical graphs of ground-penetrating radar responses of existing bridge steel bars, systematically sorts out the response characteristics, and constructs a sample data set; obtains a data set of sample data pairs through forward calculation using the finite-difference time-domain method; designs a radar intelligent identification network for bridge steel bars based on the M‑Net architecture of a fully convolutional neural network as the backbone network, and trains and optimizes the radar intelligent identification network; inverts the ground-penetrating radar data measured on the physical model to verify whether the steel bar position, dielectric constant and diameter are consistent with the actual situation. The present invention has at least the following beneficial effects: integrating multi-scale inputs into the encoding module to avoid a substantial increase in parameters; increasing the network width of the encoding module path; increasing the generalization of the BIG‑inv net network; being able to extract multi-scale information and improve the ground-penetrating radar bridge steel bar identification effect.

Description

Technical Field

[0001] The present invention relates to a radar intelligent recognition method for bridge steel bars, and in particular to the fields of bridge detection, ground penetrating radar and deep learning. Background Art

[0002] my country is a country with a large number of bridges. As an important part of transportation, bridges are indispensable for modern people's travel. Bridge projects are generally reinforced concrete structures. During the construction of concrete structures, the displacement of steel bars and the quality of steel bars are common problems, which cause many safety hazards and even safety accidents. Therefore, non-destructive testing of bridge reinforced concrete structures (the position and diameter of steel bars, etc.) is of great significance to the normal operation of bridge construction.

[0003] Ground penetrating radar technology is a non-destructive testing technology used to detect the medium (such as the earth, concrete) itself and its internal objects (such as steel bars, pipes, holes). Ground penetrating radar technology has the advantages of fast detection speed, high resolution, flexible operation, low detection cost, and wide detection range, making this technology widely used in many fields such as municipal engineering, environmental engineering, and geological engineering. Over the past decade, the rapid development of electronic components and radio technology has enabled the ground penetrating radar hardware system to have the characteristics of high acquisition accuracy, fast speed, and large data volume, while the supporting ground penetrating radar data processing technology is relatively weak, which hinders the further promotion of ground penetrating radar technology. The accuracy of ground penetrating radar data inversion depends on the experience of professionals. In addition, the amount of data collected by ground penetrating radar in actual engineering is very large, and it takes a lot of energy and time to process and interpret radar data. Therefore, improving the recognition efficiency and accuracy of steel bar diameter, position, etc. is an inevitable trend in the development of steel bar detection technology.

[0004] With the advent of the big data era, the demand for data analysis in various industries continues to increase, leading to the rapid development of machine learning. At present, many machine learning methods have been widely used to solve geophysical inverse problems and have achieved good results. Deep learning is an important branch of machine learning and has now become the core algorithm of artificial intelligence. Its essence is a composite function with universal approximation capabilities. For any nonlinear complex function, as long as there are enough neurons, the network function can arbitrarily approximate it. Its achievements in target detection, natural language processing, medical image segmentation and other fields far exceed previous related technologies.

[0005] Based on this, the present invention is based on the actual needs of intelligent identification of bridge steel bars, utilizes the unique advantages of fully convolutional neural networks in dealing with nonlinear problems, and proposes a radar intelligent identification method for bridge steel bars. Summary of the invention

[0006] In view of the above problems in the prior art, the present application proposes a radar intelligent identification method for bridge steel bars, comprising the following steps:

[0007] S1. By changing the dielectric constant and conductivity of the bridge structure, its response characteristics are obtained through forward modeling;

[0008] S2. Collect typical graphs of ground penetrating radar responses of existing bridge reinforcements, systematically sort out the response characteristics, and construct a sample data set;

[0009] S3, dividing a grid for a two-dimensional dielectric model, wherein the two-dimensional dielectric model includes steel bars of different diameters and different buried depths; designing a dielectric model of steel bars of different diameters, different buried depths and different positions in concrete, and obtaining a data set of sample data pairs by forward calculation using a finite difference time domain method; each sample data pair consists of a dielectric model and corresponding ground penetrating radar data;

[0010] S4. Design a radar intelligent recognition network for bridge steel bars based on the M-Net architecture of the fully convolutional neural network as the backbone network. Use Keras to build the radar intelligent recognition network model framework and use tensorflow as the backend.

[0011] S5. Train and optimize the radar intelligent recognition network, randomly divide the sample database into training data, test data and verification data. During the training process, all input and output values ​​are standardized to the range of [0, 1], and verification is performed after each training;

[0012] S6. Invert the measured ground penetrating radar data of the physical model to verify whether the position, dielectric constant and diameter of the steel bars are consistent with the actual situation.

[0013] In one embodiment, the dielectric constant and conductivity of the bridge structure are changed by adjusting the diameter of the steel bars, the thickness of the protective layer, and the spacing between the steel bars.

[0014] In one embodiment, the GPR data in step S3 are made to have the same size as the model data through bilinear interpolation.

[0015] In one embodiment, the radar intelligent recognition network includes a ground penetrating radar bridge steel bar multi-scale input module, an encoding module, a pyramid convolution module, a decoding module and a multi-scale output module.

[0016] In one embodiment, the ground penetrating radar bridge rebar multi-scale input module is used to construct a data pyramid input and achieve fusion of receptive fields at different levels; it uses an average pooling layer to naturally downsample the radar data and construct a multi-scale input in the encoder path.

[0017] In one embodiment, in step S4, a pyramid convolution layer is added on the basis of the M-net input layer to integrate multi-scale input into the encoding module, thereby avoiding a substantial increase in parameters, increasing the network width of the encoding module path, and increasing the generalization of the radar intelligent recognition network.

[0018] In one embodiment, the encoding module includes 8 convolutional layers and 3 pooling layers; the convolutional layer is responsible for obtaining the local features of the ground penetrating radar bridge steel bar data, and the pooling layer downsamples the data and transmits the scale-invariant features to the next layer; in order to improve the learning speed of the network while ensuring the stability of the network learning performance, a batch normalization algorithm is added after the convolutional layer, and the activation function is used to solve the problems of gradient vanishing and convergence speed during network training.

[0019] In one embodiment, after the encoding stage, the pyramid convolution module performs a pyramid convolution operation on the feature map. The pyramid convolution module includes convolution kernels of different scales, which can extract multi-scale information and improve the recognition effect of bridge steel bars by ground penetrating radar.

[0020] In one embodiment, the decoding module upsamples the reduced feature image and then performs convolution processing to improve the geometric shape of the abnormal body and compensate for the loss of details caused by the previous operation.

[0021] In one embodiment, the multi-scale output module generates a corresponding local output map for the early layers; the average of multiple local output maps is used as the final output; thereby achieving deep layer supervision, alleviating the gradient vanishing problem, and facilitating the training of the early layers.

[0022] The above technical features can be combined in various suitable ways or replaced by equivalent technical features as long as the purpose of the present invention can be achieved.

[0023] The radar intelligent identification method for bridge steel bars provided by the present invention has at least the following beneficial effects compared with the prior art:

[0024] 1) Integrate multi-scale input into the encoding module to avoid a substantial increase in parameters; 2) Increase the network width of the encoding module path; 3) Increase the generalization of the BIG-inv net network; 4) Extract multi-scale information and improve the recognition effect of bridge steel bars by ground penetrating radar; 5) Alleviate the gradient vanishing problem; 6) The intelligent recognition results are highly consistent with the model position, diameter, and dielectric constant. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The present invention will be described in more detail below based on embodiments and with reference to the accompanying drawings, wherein:

[0026] Figure 1The overall technical roadmap of the radar intelligent identification method for bridge reinforcement of the present invention is shown;

[0027] Figure 2 A schematic diagram showing the working principle of the ground penetrating radar of the present invention is shown;

[0028] Figure 3 The BIG-inv net network structure diagram of the present invention is shown;

[0029] Figure 4 A diagram showing a multi-scale input module for ground penetrating radar bridge reinforcement data of the present invention is shown;

[0030] Figure 5 shows a schematic diagram of the encoding module of the present invention;

[0031] Figure 6 A schematic diagram of a pyramid convolution module of the present invention is shown;

[0032] Figure 7 A schematic diagram of a decoding module of the present invention is shown;

[0033] Figure 8 shows a schematic diagram of a multi-scale output module of the present invention;

[0034] Figure 9(a)-9(c) The schematic diagram of intelligent identification of two steel bar models of the present invention is shown;

[0035] Figure 10(a)-10(c) The schematic diagram of intelligent identification of three steel bar models of the present invention is shown;

[0036] Fig.11 A physical model diagram of the present invention is shown;

[0037] Fig.12 A schematic diagram of intelligent recognition of measured data of the present invention is shown. DETAILED DESCRIPTION

[0038] The present invention will be further described below in conjunction with the accompanying drawings.

[0039] The general technical route of the present invention is as follows Figure 1 shown.

[0040] Ground penetrating radar uses high-frequency electromagnetic waves for detection. Figure 2As shown in the figure, the ground penetrating radar transmits high-frequency pulse electromagnetic waves underground through the transmitting antenna T. When the electromagnetic waves propagate in the underground medium, part of the radar waves will be reflected back at the place where the electromagnetic properties change. The reflected electromagnetic waves are received by the receiving antenna R, and then processed by the radar-specific software to form a two-dimensional radar data image, based on which the medium distribution law (such as medium thickness, interface, burial depth, size, shape, direction, etc. of internal buried objects or defects) can be inferred. The steel bar is a cylindrical target, which is represented by a hyperbola in the ground penetrating radar profile.

[0041] The dielectric constant anomalies in the underground half-space will cause electromagnetic field responses on the observation surface, and the electromagnetic field responses caused by different anomalies also show different characteristics. That is, they have certain spatial correlation and local existence. As a special deep learning method, the full convolutional neural network method uses convolution operators instead of matrix multiplication, focusing on learning the locality and spatiality of input images and output labels. Therefore, it is feasible to use the full convolutional neural network method to invert ground penetrating radar data.

[0042] The inversion goal of the present invention is to obtain the mapping function from the ground penetrating radar echo data to the dielectric constant of the model body through the method of full convolutional neural network. The physical parameters of the model body medium are shown in Table 1.

[0043] ε r =f(E Z ),

[0044] Where: ε r is the model dielectric constant; f is the full convolutional network mapping function; E Z Echo data.

[0045] Table 1 Physical parameters of the model medium

[0046]

[0047] Radar intelligent identification of bridge steel bars mainly includes the following steps: analysis of ground penetrating radar response characteristics of bridge steel bars, construction of sample data sets, establishment of BIG-inv net network model, network training and optimization, model testing, and instance detection.

[0048] (1) Analysis of ground penetrating radar response characteristics of bridge reinforcement

[0049] Since bridge structures are composed of a variety of media, their dielectric properties vary. By changing the dielectric constant and conductivity of concrete, as well as adjusting the steel bar diameter, protective layer thickness, steel bar spacing, etc., we can obtain its response characteristics through forward modeling, thereby understanding the propagation laws of electromagnetic waves under different conditions. In addition, we have extensively collected typical graphs of the ground penetrating radar response of existing bridge steel bars, systematically combed the response characteristics, and provided a basis for the construction of the sample data set.

[0050] (2) Sample Dataset Construction

[0051] In this paper, the two-dimensional dielectric model is divided into 256×128 grids, and the size of each grid is 0.25cm×0.25cm. The dielectric model contains steel bars of different diameters and different buried depths. The dielectric model of steel bars with different diameters, different buried depths and different positions in homogeneous concrete is designed. Through the forward calculation of the finite difference time domain method, a data set of 1152 sample data pairs is obtained. Each sample data pair consists of a dielectric model (size of 256×128) and the corresponding ground penetrating radar data (bilinear interpolation is used to make it the same size as the model data).

[0052] (3) BIG-inv net network model construction

[0053] The present invention designs a radar intelligent identification network (BIG-inv net) of bridge steel bars based on the M-Net architecture of the fully convolutional neural network as the backbone network. Figure 3 As shown in the figure, BIG-inv net includes 5 modules.

[0054] GPR bridge reinforcement data multi-scale input module Figure 4 , which is used to construct data pyramid input and realize the fusion of receptive fields at different levels. The specific implementation uses the average pooling layer to naturally downsample the radar data and construct multi-scale input in the encoder path. The present invention adds a pyramid convolution layer on the basis of the M-net input layer. The advantages of this are: 1) integrating multi-scale input into the encoding module to avoid a significant increase in parameters; 2) increasing the network width of the encoding module path; 3) increasing the generalization of the BIG-inv net network.

[0055] Encoding module see Figure 5 , this module consists of 8 convolutional layers and 3 pooling layers. The convolutional layer is responsible for obtaining the local features of the ground penetrating radar bridge steel bar data, and the pooling layer downsamples the data and transmits the scale-invariant features to the next layer. In order to ensure the stability of the network learning performance and improve the learning speed of the network, a batch normalization algorithm is added after the convolutional layer; the activation function can solve the problems of gradient disappearance and convergence speed during network training. The mathematical expression of the linear rectification function is as follows:

[0056] f ReLU (x)=max(x,0),

[0057] When the input is greater than 0, the output is equal to the input; when the input is less than 0, the output is 0.

[0058] Pyramid convolution module see Figure 6,After the encoding stage, the feature map is subjected to a pyramid ,convolution operation. This module contains convolution kernels of different scales, which can ,extract multi-scale information and improve the recognition effect of ,bridge steel bars by GPR.

[0059] Decoding module see Figure 7 ,The previous module has obtained all the anomaly information and the approximate location information, so the next step is to match these anomalies to specific pixel-level locations.,The decoding module upsamples the reduced feature image and then performs convolution processing to improve the geometric shape of the anomaly and make up the details lost by the previous ,operation.

[0060] Multi-scale output module see Figure 8 , generate a corresponding local output map for the early layers. The average of multiple local output maps is used as the final output. This enables deep layer supervision, which can alleviate the gradient vanishing problem and help the training of early layers.

[0061] (4) Network training and optimization

[0062] Network training and optimization include two parts: training data and test network. The sample database is randomly divided into training data, test data and validation data, with the corresponding ratio of 10:1:1 (training set: 920 pairs; validation set: 116 pairs; test set 116 pairs). During the training process, all input and output values ​​are standardized to the range of [0, 1]. The BIG-inv net network model framework of the present invention is constructed using Keras, and tensorflow is used as the backend. During training, the batch size is set to 10, the learning rate is 0.001, the momentum is 0.9, and the weight decay is 1e-8 Adam optimizer. A total of 60 rounds of training were performed. To prevent overfitting, verification is performed after each training. The mean square error between the theoretical model and the inversion result is set as the inversion model evaluation index, and the mean square error function expression is as follows:

[0063]

[0064] Where: n is the number of inversion parameters, m is i is the theoretical value of the dielectric constant parameter of the normalized test model, is the normalized inversion permittivity parameter.

[0065] (5) Model test

[0066] Figure 9(a) and Figure 10(a) are the echo data generated by the FDTD forward modeling of the dielectric constant model; Figure 9(b) and Figure 10(b) are the dielectric constant models; Figure 9(c) and Figure 10(c) are the recognition results of the trained neural network. The comparison table between the model recognition value and the theoretical value is as follows:

[0067] Table 2 Comparison of intelligent identification values ​​and theoretical values ​​of two steel bar models

[0068]

[0069] Table 3 Comparison of intelligent identification values ​​and theoretical values ​​of three steel bar models

[0070]

[0071]

[0072] Model tests show that the intelligent identification results are highly consistent with the model position, diameter and dielectric constant.

[0073] (6) Instance Detection

[0074] The method of the present invention is used to invert the measured ground penetrating radar data of the physical model. Fig.11 As shown in the figure, the direction of the blue arrow is the direction of the measuring line. Seven steel bars with a radius of 0.006m and eight steel bars with a radius of 0.005m are arranged vertically inside the measured object. The steel bar position parameters are shown in Table 4.

[0075] Table 4 Parameters of steel bars under test

[0076]

[0077] The acquisition equipment used in the present invention is LTD-2600 intelligent ground penetrating radar. The measuring line length is 2.56m, the time window is 10ns, and the antenna frequency is 1500MHz. The measured ground penetrating radar data and its recognition results are shown in Fig.12 It can be seen that the position, dielectric constant and diameter of the steel bars are basically consistent with the actual situation.

[0078] Although the present invention has been disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications to the technical solution of the present invention by using the methods and technical contents disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope of the technical solution of the present invention.

Claims

1. A radar intelligent identification method for bridge steel bars, characterized in that: The following steps are involved: S1. By adjusting the steel bar diameter, protective layer thickness, and steel bar spacing, the dielectric constant and conductivity of the bridge structure are changed, and its response characteristics are obtained through forward modeling; S2. Collect typical graphs of ground penetrating radar responses of existing bridge reinforcements, systematically sort out the response characteristics, and construct a sample data set; S3, dividing a grid for a two-dimensional dielectric model, wherein the two-dimensional dielectric model includes steel bars of different diameters and different buried depths; designing a dielectric model of steel bars of different diameters, different buried depths and different positions in concrete, and obtaining a data set of sample data pairs by forward calculation using a finite difference time domain method; each sample data pair consists of a dielectric model and corresponding ground penetrating radar data; and making the ground penetrating radar data and the model data the same size by bilinear interpolation; S4. Design a radar intelligent recognition network for bridge steel bars based on the M-Net architecture of the fully convolutional neural network as the backbone network. Use Keras to build the radar intelligent recognition network model framework and use tensorflow as the backend. S5. Train and optimize the radar intelligent recognition network, randomly divide the sample database into training data, test data and verification data. During the training process, all input and output values ​​are standardized to the range of [0, 1], and verification is performed after each training; S6. Invert the measured ground penetrating radar data of the physical model to verify whether the position, dielectric constant and diameter of the steel bars are consistent with the actual situation; The radar intelligent recognition network includes a ground penetrating radar bridge steel bar multi-scale input module, an encoding module, a pyramid convolution module, a decoding module and a multi-scale output module; The GPR bridge reinforcement multi-scale input module is used to construct data pyramid input and realize the fusion of receptive fields at different levels. It uses an average pooling layer to naturally downsample radar data and construct multi-scale input in the encoder path. In step S4, a pyramid convolution layer is added on the basis of the M-net input layer to integrate the multi-scale input into the encoding module, thereby avoiding a substantial increase in parameters, increasing the network width of the encoding module path, and increasing the generalization of the radar intelligent recognition network; The encoding module includes 8 convolutional layers and 3 pooling layers; the convolutional layer is responsible for obtaining the local features of the ground penetrating radar bridge steel bar data, and the pooling layer downsamples the data and transmits the scale-invariant features to the next layer; in order to ensure the stability of the network learning performance and improve the network learning speed, a batch normalization algorithm is added after the convolutional layer, and the activation function is used to solve the problems of gradient disappearance and convergence speed during network training; After the encoding stage, the pyramid convolution module performs pyramid convolution on the feature map. The pyramid convolution module contains convolution kernels of different scales, which can extract multi-scale information and improve the recognition effect of bridge steel bars by ground penetrating radar. The decoding module upsamples the reduced feature image and then performs convolution processing to improve the geometric shape of the anomaly and compensate for the loss of details caused by the previous operation; The multi-scale output module generates a corresponding local output map for the early layers; the average of multiple local output maps is used as the final output; thus achieving deep supervision, alleviating the gradient vanishing problem, and facilitating the training of early layers.

Images