Underground pipeline detection and identification method and system based on ground penetrating radar and neural network

Abstract

The application discloses a kind of based on ground penetrating radar and neural network underground pipeline detection identification method and system, first setting the transmitting antenna and receiving antenna of ground penetrating radar, radar signal is emitted through transmitting antenna, echo signal is received through receiving antenna when ground penetrating radar is detected, and underground pipeline target detection diagram P1 is generated;Then using neural network, obtain high-resolution underground pipeline target detection diagram;Front-end conditioning circuit designed in the application can amplify or attenuate signal to appropriate range according to demand.In the face of high-frequency echo signal, the application designs a high-speed data acquisition system, and a high-speed data acquisition circuit is designed based on FPGA and two pieces of domestic high-speed ADC.The application further designs a neural network framework, can realize efficient and fast multiscale feature integration, realize high-performance underground pipeline contour extraction, effectively solve the detection accuracy problem in the prior art.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of underground pipeline detection, and relates to an underground pipeline detection and identification method, system and product, in particular to an underground pipeline detection and identification method and system based on ground penetrating radar and neural networks. BACKGROUND

[0002] As the "blood vessels" of the city, underground pipelines are the "lifeline" of the normal operation of the city. With the rapid advancement of urbanization, the number and scale of urban underground pipelines are becoming larger and larger, and the construction condition is becoming more and more complex. In the construction process, due to unclear underground pipeline distribution, many difficulties are brought to urban construction and reconstruction, as well as pipeline use and maintenance.

[0003] Ground penetrating radar (GPR) is an electromagnetic technology for positioning invisible target bodies or interfaces in the underground or objects. Ground penetrating radar has become a powerful tool for geophysical exploration with its high resolution and high work efficiency. With the continuous development of signal processing technology and electronic technology, as well as the increase of engineering practice and the accumulation of experience, ground penetrating radar technology has further developed, instruments have been continuously updated, and application range has gradually expanded. Ground penetrating radar has been widely applied to the fields of construction quality control, pavement structure and material research, equipment application technology research, design data survey and collection, hidden hazard detection, engineering quality accident cause supervision and arbitration, old road evaluation and reconstruction, road structure layer thickness detection, underground pipeline detection, etc.

[0004] Ground penetrating radar uses electromagnetic waves as information carriers. Since the attenuation of electromagnetic waves in the propagation process is very fast, at the same time, the detection accuracy of underground pipelines is related to the center frequency of electromagnetic waves. The higher the center frequency of electromagnetic waves, the higher the detection accuracy, but the attenuation will also be faster, which is suitable for shallow high-precision detection, and vice versa for deep detection. Therefore, a high-power high-frequency signal is usually generated, which is transmitted to the underground through an antenna, and the echo signal is received through an antenna for analysis and processing to realize the detection of underground pipeline conditions.

[0005] Due to the working characteristics of ground penetrating radar, various kinds of clutter and noise will be received in the actual underground pipeline detection work, including electromagnetic interference caused by mobile phones, power lines, etc. The current detection image accuracy is not good, so it is urgent to study reliable ground penetrating radar echo signal processing methods and image processing methods. SUMMARY

[0006] In order to solve the technical problem that the existing ground penetrating radar cannot effectively filter electromagnetic interference in the echo signal in signal collection, resulting in poor image accuracy, the present application provides an underground pipeline detection and identification method and system based on ground penetrating radar and neural networks.

[0007] The technical scheme adopted by the method of the present application is: a ground penetrating radar and neural network based underground pipeline detection and identification method, comprising the following steps:

[0008] Step 1: Set the transmitting antenna and receiving antenna of the ground penetrating radar, transmit radar signals through the transmitting antenna, and receive echo signals during ground penetrating radar detection through the receiving antenna to generate an underground pipeline target detection map P1;

[0009] Step 2: Input the underground pipeline target detection map P1 into the neural network to obtain a high-resolution underground pipeline target detection map;

[0010] The neural network is composed of a feature extraction subnetwork and a feature fusion subnetwork;

[0011] The feature extraction subnetwork is composed of a first downsampling layer, a second downsampling layer, and a third downsampling layer connected in sequence, and is used to downsample the input P1 sequentially to obtain hierarchical image features P = {P2, P3, P4} of different sizes; P4 is a low-resolution rough detection map, denoted as P'4; the feature fusion subnetwork is composed of a first upsampling operation layer, a first fusion layer, a first feature fusion decoder, a second upsampling operation layer, a second fusion layer, a second feature fusion decoder, a third upsampling operation layer, a third fusion layer, and a third feature fusion decoder, and is used to input P'4 after upsampling processing by the first upsampling operation layer, and input the result of pixel-by-pixel addition of P'4, P3 after passing through the first fusion layer into the first feature fusion decoder for processing, and output P'3; then input P'3 after upsampling processing by the second upsampling operation layer, and input the result of pixel-by-pixel addition of P'4, P'3, P2 after passing through the second fusion layer into the second feature fusion decoder for processing, and output P'2; finally, input P'2 after upsampling processing by the third upsampling operation layer, and input the result of pixel-by-pixel addition of P'4, P'3, P'2, P1 after passing through the third fusion layer into the third feature fusion decoder for processing, to obtain a high-resolution underground pipeline target detection map.

[0012] As a preferred embodiment, the feature fusion decoder is composed of two branches arranged in parallel; the first branch is composed of a dimension reduction convolution layer, an attention convolution layer, and a classifier connected in sequence; the second branch is composed of a dimension reduction convolution layer, an attention convolution layer, an upsampling layer, and a classifier connected in sequence; wherein the output of the attention convolution layer of the first branch is added to the upsampling layer of the second branch pixel by pixel and then input into the classifier of the first branch; the high-resolution feature map is processed by the first branch to output a high-resolution detail map, the low-resolution feature map is processed by the second branch to output a low-resolution detail map; the high-resolution detail map and the low-resolution detail map are added pixel by pixel to output a high-resolution prediction map.

[0013] The technical solution adopted by the system of the present invention is: an underground pipeline detection and identification system based on ground-penetrating radar and neural network, comprising a ground-penetrating radar, an echo signal acquisition processor, and a host computer; the ground-penetrating radar comprises a main control unit, a transmitter, a transmitting antenna, a receiver, and a receiving antenna; the echo signal acquisition processor comprises an echo conditioning circuit, an ADC module, an FPGA module, and an STM32 module connected in sequence; the STM32 module is connected and communicates with the host computer;

[0014] The echo conditioning circuit is used to transform, programmably condition, and amplify the echo signal received by the ground-penetrating radar receiver. The ADC module is used to perform analog-to-digital conversion on the conditioned echo signal to obtain a parallel data stream. The FPGA module is used to control the analog-to-digital conversion process of the ADC module, and to receive the parallel data stream, process the data, and send it to the STM32 module. The STM32 module processes the received data stream and inputs it into the host computer, and sends an inverse control gain command to the FPGA module. The gain of the echo conditioning circuit is controlled by the inverse control gain command, so as to realize programmable control of the amplification or attenuation of the echo signal.

[0015] Preferably, the echo conditioning circuit includes an RF transformer, a variable gain RF amplifier, and a fully differential amplifier connected in sequence; the RF transformer is used to transform the received echo signal, and the variable gain RF amplifier is used to programmatically amplify or programmatically attenuate the transformed echo signal before inputting it to the fully differential amplifier.

[0016] Preferably, the ADC module uses two parallel analog-to-digital converters. First, it samples the echo signal output by the echo conditioning circuit using the DES bilateral sampling mode, combined with the bilateral edges of the sampling clock provided by the FPGA module, to obtain a parallel high-speed data stream. Then, it sends the parallel high-speed data stream to the FPGA module using LVDS low-voltage differential transmission.

[0017] Preferably, the FPGA module obtains two sampling clocks with a 90° phase difference by frequency multiplication and phase shifting through a PLL phase-locked loop; the FPGA module inputs sampling clocks with the same frequency but different phases to the ADC module through a high-speed data acquisition unit.

[0018] Preferably, the FPGA module includes a high-speed data acquisition unit, a data buffer unit, and a data overlay unit connected in sequence; the data overlay unit is connected to the STM32 module; the FPGA module also includes a gain control unit, which is connected to the STM32 module and the echo conditioning circuit respectively, for adjusting the gain of the echo conditioning circuit.

[0019] The data buffer unit is used to buffer the received parallel high-speed data stream and reassemble the parallel high-speed data stream into a serial stream, which is then input into the data overlay unit for data processing. The data overlay unit performs correction and noise reduction processing on the spliced ​​high-speed data stream before sending it to the STM32 module.

[0020] Preferably, the data overlay unit is used to perform zero-time correction and overlay noise reduction processing on the spliced ​​high-speed data stream.

[0021] Preferably, the zero-time correction is achieved by fixing the relative positions between the transmitting and receiving antennas and using the arrival time of the direct wave as a reference to correct the time difference.

[0022] Preferably, the superposition noise reduction process uses an average superposition noise reduction algorithm to process the high-speed data stream.

[0023] Compared with the prior art, the beneficial effects of the present invention include:

[0024] (1) The present invention designs a densely connected coarse-to-fine feature fusion decoder architecture to achieve efficient and fast multi-scale feature integration and output high-resolution underground pipeline detection maps.

[0025] (2) The present invention designs an echo conditioning circuit to condition the echo signal at the front end, and then uses FPGA and two ADCs to achieve high-speed data acquisition of 3.6GSPS, which is fast.

[0026] (3) To reduce the impact of noise, this invention designs a real-time superposition averaging noise reduction method. Simultaneously, to address signal source jitter, this invention performs zero-time correction processing on the acquired echo signals. During ground-penetrating radar operation, the FPGA performs real-time high-speed superposition averaging filtering on the received echo signal data. After a certain number of iterations, the filtered data is sent to the host computer, resulting in high detection accuracy.

[0027] (4) This invention is well applicable to various complex and changeable ground exploration environments and has high practicality. Attached Figure Description

[0028] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0029] Figure 1 This is a schematic diagram of the neural network structure according to an embodiment of the present invention;

[0030] Figure 2This is a structural diagram of the feature fusion decoder according to an embodiment of the present invention;

[0031] Figure 3 This is a system schematic diagram according to an embodiment of the present invention;

[0032] Figure 4 This is a structural diagram of the echo conditioning circuit according to an embodiment of the present invention;

[0033] Figure 5 This is a schematic diagram of the ADC module according to an embodiment of the present invention;

[0034] Figure 6 The diagram shows the experimental results of an embodiment of the present invention, where (a) is the original radar waveform output by the ground penetrating radar, and (b) is the high-resolution underground pipeline target detection map output by the neural network. Detailed Implementation

[0035] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0036] Please see Figure 1 This embodiment provides a method for detecting and identifying underground pipelines based on ground-penetrating radar and neural networks, including the following steps:

[0037] Step 1: Set up the transmitting and receiving antennas of the ground penetrating radar. Transmit radar signals through the transmitting antenna and receive echo signals from the ground penetrating radar during detection through the receiving antenna to generate an underground pipeline target detection map P1.

[0038] Step 2: Input the underground pipeline target detection map P1 into the neural network to obtain a high-resolution underground pipeline target detection map;

[0039] The neural network consists of a feature extraction subnetwork and a feature fusion subnetwork;

[0040] The feature extraction subnetwork consists of a first downsampling layer, a second downsampling layer, and a third downsampling layer connected in sequence. It is used to sequentially downsample the input P1 to obtain hierarchical image features P = {P2, P3, P4} of different sizes; where P4 is a low-resolution coarse detection map, denoted as P′4. The feature fusion subnetwork consists of a first upsampling operation layer, a first fusion layer, a first feature fusion decoder, a second upsampling operation layer, a second fusion layer, a second feature fusion decoder, a third upsampling operation layer, a third fusion layer, and a third feature fusion decoder. It is used to pass P′4 through the first upsampling operation layer... After upsampling, P′3 is added pixel-by-pixel to P′4 and P3 through the first fusion layer and then input into the first feature fusion decoder for processing, outputting P′3. Then, P′3 is upsampled through the second upsampling operation layer and added pixel-by-pixel to P′4, P′3, and P2 through the second fusion layer, then input into the second feature fusion decoder for processing, outputting P′2. Finally, P′2 is upsampled through the third upsampling operation layer and added pixel-by-pixel to P′4, P′3, P′2, and P1 through the third fusion layer, then input into the third feature fusion decoder for processing, to obtain a high-resolution underground pipeline target detection map.

[0041] This embodiment considers the input feature P. i For different numbers of feature channels, first for each input feature P i Apply dimensionality reduction convolutional layers Compress the number of feature channels c to the size of the pre-screening. This represents a convolutional layer with a kernel size of k×k and c output channels. Then, the dimensionality-reduced features are input into an attention convolutional layer to generate an attention map in the spatial dimension, suppressing background information, highlighting the baseline pipeline region, and improving model performance.

[0042] Please see Figure 2 In one embodiment, the feature fusion decoder consists of two parallel branches; the first branch consists of a sequentially connected dimensionality reduction convolutional layer, an attention convolutional layer, and a classifier; the second branch consists of a sequentially connected dimensionality reduction convolutional layer, an attention convolutional layer, an upsampling layer, and a classifier; wherein, the output of the attention convolutional layer of the first branch is added pixel-by-pixel to the upsampling layer of the second branch and then input into the classifier of the first branch; the high-resolution feature map is processed by the first branch to output a high-resolution detail map, and the low-resolution feature map is processed by the second branch to output a low-resolution detail map; the high-resolution detail map and the low-resolution detail map are added pixel-by-pixel to output a high-resolution prediction map.

[0043] The dimensionality reduction convolutional layer described in this embodiment consists of a 1×1 convolutional layer, a batch normalization layer, a ReLU activation layer, a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The attention convolutional layer consists of an average pooling layer, a first 3×3 convolutional layer, a sigmoid activation layer, a second 3×3 convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence. The original input is multiplied pixel-by-pixel by the sigmoid activation layer and then input to the second 3×3 convolutional layer for further processing.

[0044] The neural network used in this embodiment is a pre-trained network, and the training scheme adopted is an existing technical solution, which will not be described in detail here.

[0045] Please see Figure 3 This embodiment also provides an underground pipeline detection and identification system based on ground-penetrating radar and neural networks, comprising a ground-penetrating radar, an echo signal acquisition processor, and a host computer; the ground-penetrating radar consists of a main control unit, a transmitter, a transmitting antenna, a receiver, and a receiving antenna; the echo signal acquisition processor consists of an echo conditioning circuit, an ADC module, an FPGA module, and an STM32 module connected in sequence; the STM32 module is connected and communicates with the host computer;

[0046] The echo conditioning circuit is used to transform, programmably condition, and amplify the echo signal received by the ground-penetrating radar receiver. The ADC module is used to perform analog-to-digital conversion on the conditioned echo signal to obtain a parallel data stream. The FPGA module is used to control the analog-to-digital conversion process of the ADC module, and to receive the parallel data stream, process the data, and send it to the STM32 module. The STM32 module processes the received data stream and inputs it into the host computer, and sends an inverse control gain command to the FPGA module. The gain of the echo conditioning circuit is controlled by the inverse control gain command, so as to realize programmable control of the amplification or attenuation of the echo signal.

[0047] Please see Figure 4In one embodiment, the echo conditioning circuit includes a radio frequency (RF) transformer, a variable gain RF amplifier, and a fully differential amplifier connected in sequence. The RF transformer is used to transform the received echo signal, and the variable gain RF amplifier is used to programmatically amplify or attenuate the transformed echo signal before it is input to the fully differential amplifier. The echo conditioning circuit first electrically isolates the echo signal through an RF transformer. Then, due to the limited input amplitude of the high-performance ADC, it needs to pass through a variable gain RF amplifier, and the amplification or attenuation of the input signal is programmatically controlled through an FPGA module. Finally, the input signal also needs to pass through a fully differential amplifier to cancel common-mode interference before being differentially input to the ADC module. In the design of the echo conditioning circuit, impedance matching is used to achieve stable signal transmission.

[0048] In one implementation, the RF transformer can be a T3002NL model to achieve electrical isolation of the signal; the variable gain amplifier can be an ADL5201ACPZ-R7 model to perform programmable amplification or attenuation; and the fully differential amplifier can be an LMH6554 model to provide low-distortion amplification for broadband differential signals.

[0049] Please see Figure 5 In one embodiment, the ADC module employs two parallel analog-to-digital converters. First, it samples the echo signal output by the echo conditioning circuit using the DES bilateral sampling mode, combined with the bilateral edges of the sampling clock provided by the FPGA module, to obtain a parallel high-speed data stream. Then, it sends the parallel high-speed data stream to the FPGA module using LVDS low-voltage differential transmission.

[0050] In one implementation, the ADC module can employ an MXT2001 ADC, a dual-channel, high-performance ADC with an 8-bit sampling accuracy and multiple programmable acquisition and data transmission modes. In dual-channel operation, each channel can achieve a maximum sampling rate of 1.0 GSPS. It supports time-interleaved sampling, reaching a maximum of 2.0 GSPS in bilateral sampling. The on-chip ADC integrates two channels, providing four parallel high-speed data streams. Through TIADC technology, the sampling clock is phase-shifted, allowing the two ADCs to sample alternately, doubling the real-time sampling rate to 3.6 GSPS. The converted data is transmitted in parallel using LVDS, coupled with a data synchronization clock, sending the data to the FPGA module at a maximum speed of 500MHz.

[0051] In one embodiment, the FPGA module includes a high-speed data acquisition unit, a data buffer unit, and a data overlay unit connected in sequence; the data overlay unit is connected to the STM32 module; the FPGA module also includes a gain control unit, which is connected to the STM32 module and the echo conditioning circuit respectively, for adjusting the gain of the echo conditioning circuit.

[0052] The data buffer unit is used to buffer the received parallel high-speed data stream and reassemble the parallel high-speed data stream into a serial stream, which is then input into the data overlay unit for data processing. The data overlay unit performs correction and noise reduction processing on the spliced ​​high-speed data stream before sending it to the STM32 module.

[0053] The FPGA module obtains two sampling clocks with a 90° phase difference through PLL frequency multiplication and phase shifting. The FPGA module, through a high-speed data acquisition unit, inputs sampling clocks with the same frequency but different phases to the ADC module. Using sampling clocks with the same frequency but different phases controls the collaborative operation of multiple ADCs. By phase-shifting the sampling clocks and then distributing them to each sub-ADC as their sampling clocks, the same signal is acquired at different phases. The acquired data is then processed and rearranged to ultimately reconstruct the original signal.

[0054] In one implementation, the FPGA chip model of the FPGA module can be 10CL055YF484C6G.

[0055] During operation, the system in this embodiment often experiences various interferences and noises. Furthermore, due to equipment consistency issues and the temperature-dependent nature of analog circuits, the delay between the final output signal and the trigger signal may fluctuate. To reduce the impact of noise interference on detection accuracy and stability, the data needs to undergo zero-time correction and superposition noise reduction processing before being sent to the STM32 module, and then transmitted to the host computer. The STM32 module primarily serves for functional control and data interaction.

[0056] In one implementation, for zero-time correction, the approach is to fix the relative positions between the transmitting and receiving antennas and use the arrival time of the direct wave as a reference to correct the time difference. For random noise in the environment, an average superposition noise reduction algorithm is used to highlight the information of the effective reflected waveform and improve the signal-to-noise ratio.

[0057] In one implementation, zero-time correction and noise reduction can be integrated into a single module. This module includes a superposition arithmetic unit, a traversal search module, a data selector, a shift divider, and a dual-port data buffer module generated using the M9K memory resources within the FPGA. The processed data is then fed into an STM32 microcontroller, which temporarily stores the data and simultaneously controls the gain inversely. Finally, during actual detection, a normal echo signal can be observed on the host computer.

[0058] The following specific experiments further verify the effectiveness of the present invention.

[0059] This experiment was conducted on a municipal road in Wuhan. The ground-penetrating radar used was a BD-GPR-200 single-frequency ground-penetrating radar with an antenna center frequency of 200MHz, 512 sampling points, and a sampling rate of 5GHz. The detection results are shown below. Figure 6 Where (a) is the original radar waveform output by the ground-penetrating radar, and (b) is the high-resolution underground pipeline target detection map output by the neural network. Figure 6 As can be seen, this method can accurately and clearly detect underground pipelines buried at a depth of 3.8m. This method can provide an important reference for the safety of subsequent municipal pipeline design and construction.

[0060] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0061] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for detecting and identifying underground pipelines based on ground-penetrating radar and neural networks, characterized in that, Includes the following steps: Step 1: Set up the transmitting and receiving antennas of the ground-penetrating radar. Transmit radar signals through the transmitting antenna and receive the echo signals from the ground-penetrating radar during detection through the receiving antenna to generate a target map of underground pipelines. ; Step 2: Map the underground pipeline target. Inputting the data into a neural network yields a high-resolution image of underground pipeline targets. The neural network consists of a feature extraction subnetwork and a feature fusion subnetwork; The feature extraction subnetwork consists of a first downsampling layer, a second downsampling layer, and a third downsampling layer connected in sequence, used to process the input... Hierarchical image features of different sizes are obtained by sequential downsampling. Due to input features For different numbers of feature channels, first, for each input feature... Apply dimensionality reduction convolutional layers Number of feature channels Compress to the size of the pre-screening. Indicates having Core size and output channels The convolutional layer is then used; the dimensionality-reduced features are then input into the attention convolutional layer to generate an attention map in the spatial dimension; This is a low-resolution, coarse detection image, denoted as... ; The feature fusion subnetwork comprises a first upsampling operation layer, a first fusion layer, a first feature fusion decoder, a second upsampling operation layer, a second fusion layer, a second feature fusion decoder, a third upsampling operation layer, a third fusion layer, and a third feature fusion decoder, and is used to... After upsampling processing by the first upsampling operation layer, and the... , After pixel-by-pixel addition by the first fusion layer, the result is input into the first feature fusion decoder for processing, and the output is... Then After upsampling processing by the second upsampling operation layer, and then... , After pixel-by-pixel addition by the second fusion layer, the result is input into the second feature fusion decoder for processing, and the output is... Finally, After upsampling processing by the third upsampling operation layer, and then... , After pixel-by-pixel addition by the third fusion layer, the image is input into the third feature fusion decoder for processing to obtain a high-resolution underground pipeline target detection map. The feature fusion decoder consists of two parallel branches. The first branch comprises a sequentially connected dimensionality reduction convolutional layer, an attention convolutional layer, and a classifier. The second branch comprises a sequentially connected dimensionality reduction convolutional layer, an attention convolutional layer, an upsampling layer, and a classifier. The output of the attention convolutional layer in the first branch is added pixel-by-pixel to the upsampling layer in the second branch and then input into the classifier in the first branch. The high-resolution feature map is processed by the first branch to output a high-resolution detail map, and the low-resolution feature map is processed by the second branch to output a low-resolution detail map. The high-resolution detail map and the low-resolution detail map are added pixel-by-pixel to output a high-resolution prediction map. The dimensionality reduction convolutional layer consists of a 1×1 convolutional layer, a batch normalization layer, a ReLU activation layer, a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence; the attention convolutional layer consists of an average pooling layer, a first 3×3 convolutional layer, a Sigmoid activation layer, a second 3×3 convolutional layer, a batch normalization layer, and a ReLU activation layer connected in sequence; wherein, the original input is multiplied pixel-by-pixel by the Sigmoid activation layer and then input into the second 3×3 convolutional layer for subsequent processing.

2. An underground pipeline detection and identification system based on ground-penetrating radar and neural networks, used to implement the method described in claim 1; characterized in that: It consists of a ground-penetrating radar, an echo signal acquisition processor, and a host computer; the ground-penetrating radar consists of a main control unit, a transmitter, a transmitting antenna, a receiver, and a receiving antenna; the echo signal acquisition processor consists of an echo conditioning circuit, an ADC module, an FPGA module, and an STM32 module connected in sequence; the STM32 module is connected and communicates with the host computer. The echo conditioning circuit is used to transform, programmably condition, and amplify the echo signal received by the ground-penetrating radar receiver. The ADC module is used to perform analog-to-digital conversion on the conditioned echo signal to obtain a parallel data stream. The FPGA module is used to control the analog-to-digital conversion process of the ADC module, and to receive the parallel data stream, process it, and send it to the STM32 module. The STM32 module processes the received data stream and inputs it into the host computer, and sends an inverse gain control command to the FPGA module. The gain of the echo conditioning circuit is controlled by the inverse gain control command, so as to realize programmable control of the amplification or attenuation of the echo signal. The echo conditioning circuit includes an RF transformer, a variable gain RF amplifier, and a fully differential amplifier connected in sequence. The RF transformer is used to transform the received echo signal, and the variable gain RF amplifier is used to programmatically amplify or programmatically attenuate the transformed echo signal before inputting it to the fully differential amplifier. The ADC module employs two parallel-connected analog-to-digital converters. First, it samples the echo signal output from the echo conditioning circuit using the DES bilateral sampling mode, combined with the bilateral edges of the sampling clock provided by the FPGA module, to obtain a parallel high-speed data stream. Then, it sends the parallel high-speed data stream to the FPGA module using LVDS low-voltage differential transmission. The FPGA module obtains two sampling clocks with a 90° phase difference by frequency multiplication and phase shifting through a PLL phase-locked loop; the FPGA module inputs sampling clocks with the same frequency but different phases to the ADC module through a high-speed data acquisition unit. The FPGA module includes a high-speed data acquisition unit, a data buffer unit, and a data overlay unit connected in sequence; the data overlay unit is connected to the STM32 module; the FPGA module also includes a gain control unit, which is connected to the STM32 module and the echo conditioning circuit respectively, for adjusting the gain of the echo conditioning circuit. The data buffer unit is used to buffer the received parallel high-speed data streams and reassemble the parallel high-speed data streams into a serial stream, which is then input into the data overlay unit for data processing. The data overlay unit performs zero-time correction and overlay noise reduction processing on the spliced ​​high-speed data streams before sending them to the STM32 module.

3. The underground pipeline detection and identification system based on ground-penetrating radar and neural networks according to claim 2, characterized in that: The zero-time correction uses the fixed relative position between the transmitting and receiving antennas, with the arrival time of the direct wave as a reference, to correct the time difference.

4. The underground pipeline detection and identification system based on ground-penetrating radar and neural networks according to claim 2, characterized in that: The superposition noise reduction process uses an average superposition noise reduction algorithm to process the high-speed data stream.

Images