Ground penetrating radar cavity identification method, system, training method, equipment and medium

Abstract

The present disclosure relates to a method, system, training method, device and medium for identifying a hole in a ground penetrating radar. The method comprises: in the B-scan direction of the radar data, a processing window is preset, the A-scan data in the processing window is encoded, the timing information is extracted, and the encoded feature symbols are obtained; the encoded feature symbols are sequence modeled along the B-scan direction, the long-range features are captured, and the rough location information of the hole is obtained by using the OneClass anomaly detection method; the local area of ​​the rough location information of the hole is accurately extracted by using deep convolution. The present disclosure fully considers the characteristics of spatiotemporal feature extraction, long-range feature extraction and few-sample learning problems of ground penetrating radar hole identification data, and accurately identifies the specific location of the hole.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel detection, and in particular, relates to a ground penetrating radar cavity recognition method, system, training method, equipment and medium. Background Art

[0002] During the construction and service phases, the tunnel lining structure will suffer from various types of damage, such as cracks and cavities, due to the effects of gravity, load, settlement, environment and other factors. When cavities appear in the tunnel lining structure, the stress of the lining structure and the stress state of the surrounding rock will change, which will cause water seepage, instability, block loss, and detachment of the tunnel structure. In severe cases, sudden collapse may occur. The existence of cavities poses a great safety hazard to the safe operation of the tunnel. Timely discovery and identification of the location and range of tunnel lining cavities is of great significance to ensure the safe and stable service of the tunnel structure.

[0003] Geological radar is a method of detecting underground structures by emitting high-frequency electromagnetic waves underground and receiving reflected signals. The echo signal at the same location is an A-scan time-series signal. 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 representative morphological features such as hyperbolas and three-phase vibrations in the B-scan time-space signal. Using these characteristics, image analysis, mathematical modeling, machine learning, and deep learning methods can be used to interpret the location and size of the voids in radar data.

[0004] The radar B-scan spatiotemporal signal contains the A-scan temporal features and the spatial features along the B-scan direction, and there are long-range features in the B-scan direction. Only by fully mining the spatiotemporal features of the data and effectively capturing the long-range features can the hole information in the radar data be accurately interpreted.

[0005] In radar data analysis tasks, deep learning methods can transform radar B-scan data into matrix modeling (image recognition) [1] and use convolution operators for feature extraction. However, convolution operations are local receptive field feature extraction operators that cannot capture long-range features of data and are not fully suitable for radar data analysis tasks.

[0006] The analysis of radar data can also be carried out according to the sequence data modeling method, using RNN, LSTM, Transformer and other recurrent neural network algorithms for feature extraction. However, the extraction of spatiotemporal features according to the sequence data modeling method is limited, and the sequence data modeling method requires a good sequence unit encoding method. Existing model-driven encoding (such as wavelet transform method) cannot accurately perform mathematical modeling due to the irregular dielectric constant of the underground medium and various electromagnetic interferences, and the effect is limited.

[0007] Furthermore, the actual radar data of tunnel inspection contains few data samples of voids, and void identification faces a long-tail problem. If deep learning uses label supervision to train the model, the training samples will not cover the solution space enough, resulting in limited generalization ability of the model and poor robustness in actual engineering applications. Using electromagnetic simulation or generative methods to generate simulated data can alleviate this problem. However, simulated data cannot fully cover various complex scenarios and cannot really solve the problem.

[0008] It can be seen that tunnel void detection is a challenging problem, which requires full consideration of the spatiotemporal characteristics and long-range feature mining of radar data, and also needs to deal with the problem of few samples. Summary of the invention

[0009] The purpose of the present disclosure is to provide a ground penetrating radar cavity identification method, system, training method, equipment and medium in order to solve the above problems.

[0010] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0011] A ground penetrating radar cavity identification method comprises the following steps:

[0012] In the B-scan direction of the radar data, a processing window is preset, the A-scan data within the processing window is encoded, the timing information is extracted, and the encoded characteristic symbols are obtained;

[0013] Sequence modeling is performed on the encoded feature symbols along the B-scan direction to capture long-range features, and a OneClass anomaly detection method is used to obtain rough location information of the void;

[0014] The local area of ​​the rough hole position information is subjected to deep convolution to accurately extract the hole position information.

[0015] As a further optimization scheme of the present disclosure, in the B-scan direction of the radar data, a processing window is preset, the A-scan data in the processing window is encoded, the timing information is extracted, and the encoded characteristic symbols are obtained specifically including:

[0016] Extract each A-scan data , increase w weight parameter and offset , the neural network adaptively confirms the participation of each A-scan data , the formula is as follows:

[0017] ;

[0018] Add dimension expression to each A-scan and increase participation Each value in Through full connection mapping to a high-dimensional space vector representation of size n, the formula is as follows:

[0019] ;

[0020] in, for Mapping result expression;

[0021] The high-dimensional space vector Perform self-attention mechanism processing to obtain self-attention data;

[0022] The self-attention data is mapped to a feature vector of size 1xV to obtain the encoded feature symbol.

[0023] As a further optimization scheme of the present disclosure, the encoded feature symbols are sequence modeled along the B-scan direction to capture long-range features, and the OneClass anomaly detection method is used to obtain rough location information of the void, which specifically includes:

[0024] Performing sequence modeling on the encoded feature symbols along the B-scan direction to extract sequence features, including long-range features;

[0025] Performing dense sampling and sparse sampling on the extracted sequence features to obtain sampling data, and mapping the sampling data and the sequence features using the OneClass anomaly detection method to obtain a probability density distribution;

[0026] The rough location information of the hole is summarized based on the probability density distribution.

[0027] As a further optimization scheme of the present disclosure, the local area of ​​the rough hole position information is subjected to deep convolution to accurately extract the hole position information, specifically including:

[0028] A transposed convolution operation is performed on the rough position information of the hole to obtain a spatial recognition probability map. The shape of the spatial recognition probability map is the same as the shape of the input data. Each value of the spatial recognition probability map represents the hole probability of the corresponding position, thereby calculating the hole position and size.

[0029] A ground penetrating radar cavity identification system, comprising:

[0030] The encoding module is used to preset a processing window in the B-scan direction of the radar data, encode the A-scan data in the processing window, extract the timing information, and obtain the encoded feature symbols;

[0031] A symbol sequence anomaly detection module is used to perform sequence modeling on the encoded feature symbols along the B-scan direction, capture long-range features, and use the OneClass anomaly detection method to obtain rough location information of the void;

[0032] The local space positioning detection module is used to use deep convolution to accurately extract the hole position information from the local area of ​​the rough hole position information.

[0033] As a further optimization scheme of the present disclosure, the encoding module includes a window selection module, a feature dimension enhancement module, a self-attention module, and an output module;

[0034] The window selection module extracts each A-scan data , increase w weight parameter and offset , the neural network adaptively confirms the participation of each A-scan data , the formula is as follows:

[0035] ;

[0036] The feature dimension-enhancing module adds dimension expression to each A-scan, Each value in Through full connection mapping to a high-dimensional space vector representation of size n, the formula is as follows:

[0037] ;

[0038] in, for Mapping result expression;

[0039] The self-attention module converts the high-dimensional space vector Perform self-attention mechanism processing to obtain self-attention data;

[0040] The output module maps the self-attention data to a feature vector of size 1xV to obtain an encoded feature symbol.

[0041] As a further optimization solution of the present disclosure, the symbol sequence anomaly detection module includes a sequence feature extraction module and a normalized flow module;

[0042] The sequence feature extraction module performs sequence modeling on the encoded feature symbols along the B-scan direction and then extracts sequence features, including long-range features;

[0043] The normalization flow module performs dense sampling and sparse sampling on the extracted sequence features to obtain sampling data, and maps the sampling data and the sequence features using the OneClass anomaly detection method to obtain a probability density distribution; based on the probability density distribution, the rough location information of the void is summarized.

[0044] As a further optimization scheme of the present disclosure, the local space positioning detection module includes a positioning convolution block and a transposed convolution;

[0045] The positioning convolution block and the transposed convolution operate on the rough position information of the hole to obtain a spatial recognition probability map. The shape of the spatial recognition probability map is the same as the shape of the input data. Each value of the spatial recognition probability map represents the hole probability of the corresponding position, thereby calculating the hole position and size.

[0046] A training method for a ground penetrating radar cavity recognition system comprises the following steps:

[0047] The training methods of the encoding module include:

[0048] For the radar B-scan data along the survey line direction, use a sliding window of size M to intercept the radar B-scan data block and record the intercepted results. , where a single data block The shape is , where N represents the number of A-scan data sampling points, and M represents the number of intercepted A-scan data;

[0049] right Perform data enhancement operations, including up and down translation, random noise and drop operations, to obtain enhanced data ;

[0050] The encoding network model adopts self-supervised contrastive learning training method, taking training samples from input_b and input_b+ as the input of the encoder network, where bi and its own enhanced data b+i constitute positive samples, and bi and other samples constitute negative samples;

[0051] Mapping vectors to output modules , the loss function formula of the encoding module is as follows:

[0052] ;

[0053] in, To adjust the parameters;

[0054] The symbol sequence anomaly detection module training method includes:

[0055] The one-class architecture for anomaly detection is adopted. The model training only has positive samples and does not contain holes. The samples are generated uniformly by the encoding model and then cut into L*V size data blocks. The data blocks use random cutting, drop and random self-copying to insert data. The symbol sequence anomaly detection model adopts the semi-supervised probability density distribution training method. The output module is mapped to vectors. ,The loss function formula of the symbol sequence anomaly detection module is as follows:

[0056] ;

[0057] The local space positioning detection module training method includes:

[0058] The dataset uses data samples that only contain holes, and uses random cropping, rotation, and random noise data enhancement methods for the training samples; the local space positioning detection model uses supervised learning training, and the loss function is as follows:

[0059] ;

[0060] Among them, p(x) is the model output and y is the true label.

[0061] An electronic device comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0062] Memory, for storing computer programs;

[0063] The processor is used to implement the ground penetrating radar cavity identification method when executing the program stored in the memory.

[0064] A computer-readable storage medium stores a computer program, wherein the computer program implements a ground penetrating radar cavity identification method when executed by a processor.

[0065] The beneficial effects of the present disclosure are:

[0066] The present invention fully considers the characteristics of spatiotemporal feature extraction, long-range feature extraction and few-sample learning problems of ground penetrating radar cavity identification data to accurately identify the specific location of the cavity. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of the method disclosed herein;

[0068] Figure 2 is a system identification flow chart in an embodiment of the present disclosure;

[0069] Figure 3 is a system training flow chart in an embodiment of the present disclosure;

[0070] Figure 4 is a network structure diagram of an encoding module in an embodiment of the present disclosure;

[0071] Figure 5 is a network structure diagram of a symbol sequence anomaly detection module in an embodiment of the present disclosure;

[0072] Figure 6 is a network structure diagram of a normalized flow module in an embodiment of the present disclosure;

[0073] Figure 7 is a network structure diagram of a fused one-dimensional convolution module in an embodiment of the present disclosure;

[0074] Figure 8 is a network structure diagram of a local space positioning detection module in an embodiment of the present disclosure;

[0075] Fig. 9 is a system structure block diagram in an embodiment of the present disclosure;

[0076] Fig.10 It is a block diagram of the device structure in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0077] The present application is further described in detail below in conjunction with the accompanying drawings. It is necessary to point out here that the following specific implementation methods are only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Technical personnel in this field can make some non-essential improvements and adjustments to the present application based on the above application content.

[0078] like Figure 1 As shown, a ground penetrating radar cavity identification method comprises the following steps:

[0079] In the B-scan direction of the radar data, a processing window is preset, the A-scan data within the processing window is encoded, the timing information is extracted, and the encoded characteristic symbols are obtained;

[0080] Sequence modeling is performed on the encoded feature symbols along the B-scan direction to capture long-range features, and a OneClass anomaly detection method is used to obtain rough location information of the void;

[0081] The local area of ​​the rough position information of the hole is subjected to deep convolution to accurately extract spatial information.

[0082] In the B-scan direction of the radar data, a processing window is preset, the A-scan data in the processing window is encoded, the timing information is extracted, and the encoded characteristic symbols are obtained, including:

[0083] Extract each A-scan data , increase w weight parameter and offset , the neural network adaptively confirms the participation of each A-scan data , the formula is as follows:

[0084] ;

[0085] Add dimension expression to each A-scan and increase participation Each value in Through full connection mapping to a high-dimensional space vector representation of size n, the formula is as follows:

[0086] ;

[0087] in, for Mapping result expression;

[0088] The high-dimensional space vector Perform self-attention mechanism processing to obtain self-attention data;

[0089] The self-attention data is mapped to a feature vector of size 1xV to obtain the encoded feature symbol.

[0090] The encoded feature symbols are sequence modeled along the B-scan direction to capture long-range features. The OneClass anomaly detection method is used to obtain rough location information of the void, including:

[0091] Performing sequence modeling on the encoded feature symbols along the B-scan direction to extract sequence features, including long-range features;

[0092] Performing dense sampling and sparse sampling on the extracted sequence features to obtain sampling data, and mapping the sampling data and the sequence features using the OneClass anomaly detection method to obtain a probability density distribution;

[0093] The rough location information of the hole is summarized based on the probability density distribution.

[0094] The local area of ​​the rough hole position information is subjected to deep convolution to accurately extract the hole position information, specifically including:

[0095] A transposed convolution operation is performed on the rough position information of the hole to obtain a spatial recognition probability map. The shape of the spatial recognition probability map is the same as the shape of the input data. Each value of the spatial recognition probability map represents the hole probability of the corresponding position, thereby calculating the hole position and size.

[0096] like Fig. 9As shown, an embodiment of the present disclosure provides a ground penetrating radar cavity identification system, including: an encoding module, a symbol sequence anomaly detection module, and a local space positioning detection module.

[0097] The encoding module solves the problem of A-scan data encoding and representation.

[0098] The symbol sequence anomaly detection module is used to roughly locate the position of the hole in the B-scan direction.

[0099] The local spatial positioning detection module is used to accurately identify the specific location of the cavity in the coarsely positioned B-scan local area.

[0100] The encoding module can effectively extract the time series features of radar data in the A-scan direction and adopts a self-supervised learning training method.

[0101] The symbol sequence anomaly detection module can effectively extract the long-range features of radar data in the B-scan direction, and the training adopts the classic anomaly detection one-class semi-supervised learning scheme.

[0102] The local space positioning detection module is trained using supervised learning methods;

[0103] This public identification process is shown in Figure 2 , the radar B-scan data is input into the encoding module, and after encoding, the encoded characteristic symbols are formed, which are input into the symbol sequence anomaly detection module to obtain the rough location information of the hole in the radar data in the survey line direction; the local space positioning detection module only recognizes the data of the rough location information of the hole, and finally confirms the specific location information of the hole.

[0104] The identification steps of this scheme are as follows:

[0105] Step 1: For the radar B-scan data (shape MxN), take each A-scan as the center, and sequentially take WxN data blocks according to the sliding window and input them into the encoding module. After all data blocks are calculated by the encoding module, the encoded feature symbols are formed.

[0106] Step 2: Input the encoded feature symbols formed in step 1 into the symbol sequence anomaly detection module, which can map normal data to the same probability density distribution. The output value of the symbol sequence anomaly detection module is a probability density estimate corresponding to the abnormal judgment of each symbol. In this way, the encoded feature symbols are classified as normal or abnormal, the distribution of abnormal symbols can be obtained, and the rough position information of the void can be summarized, which is equivalent to the rough positioning of the void in the B-scan direction.

[0107] Step 3: According to the rough location information of the hole located in step 2, the corresponding data area block is intercepted and input into the local space positioning detection module. The local space positioning detection module is mainly composed of convolution and transposed convolution operators, and can accept data of any size. After the local space positioning detection module is operated, the spatial recognition probability map is obtained to accurately locate the hole location information.

[0108] The specific identification steps are as follows:

[0109] Assume that the network structure of the encoding module in step 1 is as shown in the figure. The network architecture includes a window selection module, a feature dimension upgrade module, a self-attention module, and an output module. Figure 4 shown.

[0110] The window selection module will extract each A-scan data , increase w weight parameter and offset , the neural network adaptively confirms the participation degree of each A-scan data.

[0111] (1);

[0112] The feature dimension-increasing module adds dimension expression to each A-scan and calculates the result of the window selection module. Each value in Through full connection mapping to a high-dimensional space vector representation of size n, for The mapping result is expressed and then added as the input of the self-attention module.

[0113] (2);

[0114] Self-attention module, data passes through this module to obtain the output of the self-attention mechanism.

[0115] The output module maps the data to a feature vector of size 1xV.

[0116] The entire encoding process is that the radar B-scan data (shape MxN) is centered on each A-scan, and the data blocks of size WxN are taken in sequence according to the sliding window. The data blocks are mapped to feature vectors of size 1xV through the window selection module, feature dimension upgrade module, self-attention module, and output module. All data blocks perform the same encoding operation to form encoded feature symbols. The length remains unchanged after encoding, and does not affect the sequence annotation label.

[0117] Assume that the network structure of the symbol sequence anomaly detection module in step 2 includes a sequence feature extraction module and a normalized flow module. Figure 5 .

[0118] Among them, the sequence feature extraction module uses the transformer self-attention architecture to extract sequence features from the data.

[0119] Normalized flow module, using a multi-scale fusion network structure, see Figure 6 The normalized flow module performs dense sampling and sparse sampling on the input data. Dense sampling is generated by the self-replication insertion method, and the data length is doubled. Sparse sampling uses the interval extraction method, and the length is 1 / 2 of the original data. The three copies of data are input into the normalized flow module at the same time. The calculation formula refers to (3) and (4), where and is the input data, and is the output of the fused one-dimensional convolution module, To adjust the parameters.

[0120] (3);

[0121] (4);

[0122] The normalized flow module uses a fused one-dimensional convolution module, such as Figure 7 As shown in Figure 2. By using the fused one-dimensional convolution module, the multi-scale merging operation of the data is completed. ,and , do the splicing operation, and then map it to a probability density distribution through the square operation and the mean operation. This distribution is equivalent to the recognition of the rough position information of the hole in the B-scan direction.

[0123] The encoded data is mapped to the same probability density distribution through the sequence feature extraction module and the normalized flow module. Each probability density estimate corresponds to the abnormal judgment of each A-scan. Thus, a rough location information of the hole is obtained, which is equivalent to the rough positioning of the hole in the B-scan direction.

[0124] Assume the network structure of the local space positioning detection module in step 3, such as Figure 8 As shown in the figure. The model architecture is mainly composed of localization convolution blocks and transposed convolutions. The input data is processed by several localization convolution blocks and transposed convolutions to obtain the output result, which is a matrix probability map. The shape of the probability map is the same as the input data shape. Each value in the probability map represents the hole probability at the corresponding position, from which the hole position and size can be calculated. The local space positioning detection module does not contain operators that have size requirements such as full connection, so it can accept data input of any size.

[0125] The identification of holes by local spatial positioning detection depends on the rough hole location information of the radar data in the survey line direction identified in step 2. The rough hole location information can be cropped as needed without the need for size conversion. The cropped data is processed by the model to obtain a matrix probability map, which is then post-processed to locate the hole position and size.

[0126] The implementation process of the functions and effects of each module in the above system is specifically described in the implementation process of the corresponding steps in the above method, which will not be repeated here.

[0127] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The system embodiment described above is only schematic, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without paying any creative work.

[0128] In the above embodiments, any number of all modules can be combined into one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. At least one of all modules can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware and firmware or in a suitable combination of any of them. Alternatively, at least one of all modules can be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding function can be executed.

[0129] The ground penetrating radar cavity recognition system training method includes the following steps:

[0130] The encoding module, symbol sequence anomaly detection module and local space positioning detection module of the present disclosure need to be trained independently. The training process of each module is shown in Figure 3 As shown in the figure, the training of the symbol sequence anomaly detection module depends on the encoding of radar data by the encoding module, and the training of the encoding module must be completed first.

[0131] The first encoding module training method uses a sliding window of size M to intercept B-scan data blocks along the survey line direction for radar data B-scan data, and records the intercepted results. , 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 .

[0132] The encoding network model adopts self-supervised contrastive learning training method. Training samples are taken from input_b and input_b+ as the input of the encoder network, where bi and its own enhanced data b+i constitute positive samples, and bi and other samples constitute negative samples.

[0133] Mapping vectors to output modules , the loss function formula of the encoding module is as follows:

[0134] (5);

[0135] in To adjust the parameters.

[0136] The second symbol sequence anomaly detection module training method uses a one-class anomaly detection architecture due to insufficient hole data samples. Therefore, the model training only has positive samples and does not contain holes. The samples are generated uniformly by the encoding model and then cut into LxV size data blocks. This data block uses random cutting, drop and random self-copy insertion data enhancement methods. The symbol sequence anomaly detection module uses a semi-supervised probability density distribution training method. The output module maps the vector , the loss function formula of the symbol sequence anomaly detection module is as follows:

[0137] (6);

[0138] The third local space positioning detection module training method uses a data set that contains only void data samples. The training samples are enhanced using random cropping, rotation, and random noise data. The local space positioning detection model uses supervised learning training, and the loss function is as follows:

[0139] (7);

[0140] Where p(x) is the model output and y is the true label.

[0141] The present disclosure uses about 151G of tunnel radar data, of which about 149G does not include void data and about 1G includes void data.

[0142] The encoding module of this scheme uses about 151G of unselected tunnel radar data for self-supervised comparative learning training.

[0143] The symbol sequence anomaly detection module of this scheme uses about 149G of radar data without holes for semi-supervised training.

[0144] The local spatial positioning detection module of this scheme uses about 900M hole data for supervised training and about 200M hole data as a validation set.

[0145] The recognition results are shown in Table 1, where mIOU represents the average intersection-over-union ratio of hole recognition, and FPS represents the average number of recognitions per second. All indicators in the table are obtained on the NVIDIA Tesla V100 computing platform.

[0146] Table 1 Void Identification

[0147]

[0148] See also Fig.10 , an electronic device provided by an embodiment of the present disclosure includes a processor 1110, a communication interface 1120, a memory 1130 and a communication bus 1140, wherein the processor 1110, the communication interface 1120, and the memory 1130 communicate with each other through the communication bus 1140;

[0149] Memory 1130, for storing computer programs;

[0150] The processor 1110 is used to implement the ground penetrating radar cavity identification method as shown below when executing the program stored in the memory 1130.

[0151] The communication bus 1140 may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus 1140 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0152] The communication interface 1120 is used for communication between the above electronic device and other devices.

[0153] The memory 1130 may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory 1130 may also be at least one storage device located away from the processor 1110.

[0154] The above-mentioned processor 1110 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.

[0155] The embodiment of the present disclosure further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the ground penetrating radar cavity identification method described above is implemented.

[0156] The computer-readable storage medium may be included in the device / apparatus described in the above embodiment; or it may exist independently without being assembled into the device / apparatus. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the ground penetrating radar cavity identification method according to the embodiment of the present disclosure is implemented.

[0157] According to an embodiment of the present disclosure, a computer-readable storage medium may be a non-volatile computer-readable storage medium, for example, may include but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, apparatus, or device.

[0158] The above-mentioned embodiments only express several implementation methods of the present disclosure, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present disclosure. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present disclosure, and these all belong to the protection scope of the present disclosure.

Claims

1. A ground penetrating radar cavity identification method, characterized in that: The following steps are involved: In the B-scan direction of the radar data, a processing window is preset, the A-scan data within the processing window is encoded, the timing information is extracted, and the encoded characteristic symbols are obtained; The encoded feature symbols are sequence-modeled along the B-scan direction to capture long-range features, and the rough location information of the void is obtained by using the OneClass anomaly detection method, specifically including: extracting sequence features, including long-range features, after sequence-modeling the encoded feature symbols along the B-scan direction; Dense sampling and sparse sampling are performed on the extracted sequence features to obtain sampling data, and the sampling data and the sequence features are mapped using the OneClass anomaly detection method to obtain a probability density distribution; based on the probability density distribution, rough location information of the void is summarized; The local area of ​​the rough hole position information is subjected to deep convolution to accurately extract the hole position information.

2. A ground penetrating radar cavity identification method according to claim 1, characterized in that: In the B-scan direction of the radar data, a processing window is preset, the A-scan data in the processing window is encoded, the timing information is extracted, and the encoded characteristic symbols are obtained, including: Extract each A-scan data , increase w weight parameter and offset , the neural network adaptively confirms the participation of each A-scan data , the formula is as follows: ； Add dimension expression to each A-scan and increase participation Each value in Through full connection mapping to a high-dimensional space vector representation of size n, the formula is as follows: ； in, for Mapping result expression; The high-dimensional space vector Perform self-attention mechanism processing to obtain self-attention data; The self-attention data is mapped to a feature vector of size 1xV to obtain the encoded feature symbol.

3. The method for identifying a cavity using ground penetrating radar according to claim 1, characterized in that: The local area of ​​the rough hole position information is subjected to deep convolution to accurately extract the hole position information, specifically including: A transposed convolution operation is performed on the rough position information of the hole to obtain a spatial recognition probability map. The shape of the spatial recognition probability map is the same as the shape of the input data. Each value of the spatial recognition probability map represents the hole probability of the corresponding position, thereby calculating the hole position and size.

4. A ground penetrating radar cavity identification system, characterized in that: include: The encoding module is used to preset a processing window in the B-scan direction of the radar data, encode the A-scan data in the processing window, extract the timing information, and obtain the encoded feature symbols; A symbol sequence anomaly detection module is used to perform sequence modeling on the encoded feature symbols along the B-scan direction, capture long-range features, and use the OneClass anomaly detection method to obtain rough location information of the void; the symbol sequence anomaly detection module includes a sequence feature extraction module and a normalization flow module; the sequence feature extraction module performs sequence modeling on the encoded feature symbols along the B-scan direction and then extracts sequence features, including long-range features; the normalization flow module performs dense sampling and sparse sampling on the extracted sequence features to obtain sampling data, and maps the sampling data and the sequence features using the OneClass anomaly detection method to obtain a probability density distribution; based on the probability density distribution, the rough location information of the void is summarized; The local space positioning detection module is used to use deep convolution to accurately extract the hole position information from the local area of ​​the rough hole position information.

5. A ground penetrating radar cavity identification system according to claim 4, characterized in that: The encoding module includes a window selection module, a feature dimension enhancement module, a self-attention module, and an output module; The window selection module extracts each A-scan data , increase w weight parameter and offset , the neural network adaptively confirms the participation of each A-scan data , the formula is as follows: ； The feature dimension-enhancing module adds dimension expression to each A-scan, Each value in Through full connection mapping to a high-dimensional space vector representation of size n, the formula is as follows: ； in, for Mapping result expression; The self-attention module converts the high-dimensional space vector Perform self-attention mechanism processing to obtain self-attention data; The output module maps the self-attention data to a feature vector of size 1xV to obtain an encoded feature symbol.

6. A ground penetrating radar cavity identification system according to claim 4, characterized in that: The local space positioning detection module includes a positioning convolution block and a transposed convolution; The positioning convolution block and the transposed convolution operate on the rough position information of the hole to obtain a spatial recognition probability map. The shape of the spatial recognition probability map is the same as the shape of the input data. Each value of the spatial recognition probability map represents the hole probability of the corresponding position, thereby calculating the hole position and size.

7. A training method for a ground penetrating radar cavity recognition system, characterized in that: The following steps are involved: The training methods of the encoding module include: For the radar B-scan data along the survey line direction, use a sliding window of size M to intercept the radar B-scan data block and record the intercepted results. , where a single data block The shape is , where N represents the number of A-scan data sampling points, and M represents the number of intercepted A-scan data; right Perform data enhancement operations, including up and down translation, random noise and drop operations, to obtain enhanced data ; The encoding network model adopts self-supervised contrastive learning training method, taking training samples from input_b and input_b+ as the input of the encoder network, where bi and its own enhanced data b+i constitute positive samples, and bi and other samples constitute negative samples; Mapping vectors to output modules , the loss function formula of the encoding module is as follows: ； in, To adjust the parameters; The symbol sequence anomaly detection module training method includes: The one-class architecture for anomaly detection is adopted. The model training only has positive samples and does not contain holes. The samples are generated uniformly by the encoding model and then cut into L*V size data blocks. The data blocks use random cutting, drop and random self-copying to insert data. The symbol sequence anomaly detection model adopts the semi-supervised probability density distribution training method. The output module is mapped to vectors. ,The loss function formula of the symbol sequence anomaly detection module is as follows: ； The local space positioning detection module training method includes: The dataset uses data samples that only contain holes, and uses random cropping, rotation, and random noise data enhancement methods for the training samples; the local space positioning detection model uses supervised learning training, and the loss function is as follows: ； Among them, p(x) is the model output and y is the true label.

8. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory, for storing computer programs; The processor is used to implement the ground penetrating radar cavity identification method described in any one of claims 1 to 3 when executing the program stored in the memory.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the ground penetrating radar cavity identification method according to any one of claims 1 to 3 is implemented.

Images