A method for identifying underground cavity targets using ground penetrating radar based on BP neural network

Abstract

The present invention provides a method for identifying underground cavity targets of ground penetrating radar based on BP neural network, including: using robust principal component analysis to suppress clutter on ground penetrating radar B-Scan images; extracting time domain, frequency domain and wavelet domain features of A-Scan signals; using BP neural network to identify A-Scan signals as a whole to obtain the target horizontal direction area; segmenting the A-Scan signals in the target horizontal direction area, extracting the time domain, frequency domain and wavelet domain features of each signal segment; using BP neural network to identify each signal segment to obtain the target vertical direction area in each A-Scan signal; determining the target position according to the target horizontal direction and vertical direction areas. The present invention has the beneficial effects of: while ensuring the target recognition accuracy, it can improve the clutter suppression effect and effectively improve the target recognition efficiency, and is suitable for the detection and positioning of underground cavity targets on roads.

Description

Technical Field

[0001] The invention relates to the field of radar signal processing and target recognition, and in particular to a method for identifying underground cavity targets of a ground penetrating radar based on a BP neural network. Background Art

[0002] With the continuous development of my country's economy, the level of urbanization has gradually increased, and the pressure of urban traffic has also increased. Detecting and repairing underground road diseases is a key issue in road maintenance. Underground cavities are the most common road diseases, which can easily lead to ground collapse, which not only has a serious impact on urban traffic, but also endangers people's lives and property. Therefore, the detection and identification of underground cavities is of great significance to ensure the safety of urban roads.

[0003] Ground penetrating radar has the advantages of fast detection speed, continuous detection process, high resolution, easy operation, low detection cost, wide detection range and intuitive profile diagram, and is increasingly used in underground cavity detection in roads. Traditional ground penetrating radar profile interpretation mainly relies on the experience of staff, which is prone to errors and low efficiency in the interpretation process. In order to improve the detection accuracy and efficiency, it is necessary to study the automatic recognition technology of targets.

[0004] At present, the ground penetrating radar target recognition methods mainly include machine learning methods (support vector machines and neural networks) and deep learning methods (convolutional neural networks). The machine learning method needs to extract features before recognition, and the network structure is simple and the running speed is fast; the deep learning method directly uses the deep network to extract features, the network structure is complex, and a large number of labeled samples are required for training, and the recognition speed is slow. Therefore, from the perspective of recognition efficiency, the machine learning method has more advantages. Among the machine learning methods, from the perspective of feature extraction, it can be mainly divided into two categories based on A-Scan signals (single-channel signals) and B-Scan images. The feature extraction method based on A-Scan signals extracts features from the one-dimensional data of the ground penetrating radar, and the calculation is relatively simple; the feature extraction method based on B-Scan images mainly extracts features from two-dimensional images, and the calculation is relatively complex. Therefore, the feature extraction method based on A-Scan signals has a smaller amount of calculation.

[0005] In the recognition of cavity targets in ground penetrating radar, the recognition methods based on A-Scan signal feature extraction can be divided into two categories: one is to extract the overall features of an A-Scan signal and use a classifier to identify whether the signal contains a cavity target. This method has high recognition efficiency, but can only identify the horizontal position of the target and is used for preliminary judgment of the cavity; the other is to extract the features of the A-Scan segmented signal and use a classifier to identify whether the segmented signal is a cavity target echo. This method can identify the target position and is useful for judgment of the cavity, but for the entire B-Scan image, the number of segmented signals is large, and the feature extraction calculation amount is also large, which affects the recognition efficiency. In actual road underground cavity detection, the number of A-Scan signals containing cavity targets is small, and it is not necessary to recognize all A-Scan segmented signals. Therefore, how to reduce the recognition of invalid segmented signals and improve recognition efficiency is of great significance to improving the cavity detection performance of ground penetrating radar. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a method for identifying underground cavity targets of ground penetrating radar based on BP neural network. The method for identifying underground cavity targets of ground penetrating radar uses robust principal component analysis method to decompose the ground penetrating radar B-Scan image to better distinguish clutter and target components, which can improve the clutter suppression effect; extract multidimensional combination features in time domain, frequency domain and wavelet domain, which can better reflect the difference between target signal and non-target signal; use BP neural network to perform overall recognition and then segmented recognition on A-Scan signal to reduce the amount of calculation, and by combining overall recognition and segmented recognition, while ensuring the accuracy of target recognition, the recognition efficiency can be effectively improved, which is suitable for the detection and positioning of underground cavity targets on roads. The method for identifying underground cavity targets of ground penetrating radar mainly includes the following steps:

[0007] S1: Acquire the ground penetrating radar B-Scan image and process the ground penetrating radar B-Scan image using the robust principal component analysis method to suppress clutter signals;

[0008] S2: Extract features from each A-Scan signal in the processed B-Scan image to obtain features in the time domain, frequency domain, and wavelet domain respectively;

[0009] S3: Using BP neural network to identify each A-Scan signal in the B-Scan image after clutter suppression, and obtain the target horizontal area;

[0010] S4: segmenting the A-Scan signal in the target horizontal area, and extracting the time domain, frequency domain and wavelet domain features of each segment of the signal;

[0011] S5: According to the time domain, frequency domain and wavelet domain features of each segment of the signal obtained in step S4, the segmented signals in the A-Scan are identified using a BP neural network to obtain the target vertical direction area in the A-Scan signal;

[0012] S6: Determine the target position according to the target horizontal direction area and the target vertical direction area.

[0013] Furthermore, the steps of using robust principal component analysis to suppress clutter signals in ground penetrating radar B-Scan images are as follows:

[0014] (1) Obtain the ground penetrating radar B-Scan image X∈R M×N , where N is the number of A-Scan signals in the B-Scan image, M is the number of sampling points of each A-Scan signal, X = [x 1 (m),x 2 (m),…,x N (m)],x i (m) = (x i (1),x i (2),…,x i (M)) T is the i-th A-Scan signal, m represents the sample point number in each A-Scan signal, 1≤m≤M;

[0015] (2) Each A-Scan signal in the ground penetrating radar B-Scan image X is processed to remove the DC offset. The processing process is as follows:

[0016]

[0017] in, It represents the mean value of each A-Scan signal data in the ground penetrating radar B-Scan image X.

[0018] (3) The robust principal component analysis method is used to process the ground penetrating radar B-scan image after removing the DC offset to obtain a low-rank matrix G and a sparse matrix S. The sparse matrix S is selected as the target echo signal to obtain the B-scan image X1 after clutter suppression.

[0019] Furthermore, the time domain, frequency domain and wavelet domain features obtained in step S2 are as follows:

[0020] (1) Time domain characteristics

[0021] The time domain features include 3 features: mean MEAN i , standard deviation STD i and RMS i , the calculation process is as follows:

[0022]

[0023]

[0024]

[0025] Among them, x1 i (m) is the i-th A-Scan signal data in the B-Scan image after clutter suppression;

[0026] (2) Frequency domain characteristics

[0027] The frequency domain feature is the signal energy spectrum, which is calculated as follows:

[0028]

[0029] Among them, k1 is the frequency point number;

[0030] (3) Wavelet domain characteristics

[0031] The wavelet domain feature is the wavelet packet energy spectrum T i , the wavelet packet energy spectrum is defined as the normalized wavelet packet node energy:

[0032]

[0033] Wavelet packet node energy e i,l,j for:

[0034]

[0035] Among them, i is the track number, For x1 i (m) is the wavelet packet decomposition coefficient, l is the number of decomposition levels, j is the frequency band, and k2 is the length of the decomposition coefficient.

[0036] Furthermore, the steps of obtaining the target horizontal area are as follows:

[0037] (1) Multiple A-Scan signals with and without hole targets are selected from the B-Scan image after clutter suppression, and features are extracted to construct a sample set, where each sample is a 23-dimensional feature vector;

[0038] (2) selecting part of the data from the sample set as a training set, using the training set to train the BP neural network to obtain a network model NET1, and using NET1 to identify the sample set;

[0039] (3) For K A-Scan signals identified as containing targets According to its horizontal position i k, the initial target horizontal direction region H1 = {i k , 1 ≤ k ≤ K};

[0040] (4) Perform fusion and deletion processing on the initial target horizontal direction region H1, and the specific process is as follows:

[0041] 1) Fusion processing

[0042] Suppose and are two adjacent A-Scan signals containing the target. If their trace difference 1 < |i k - i k+1 | ≤ id, then all A-Scan signals x1 i (m) between these adjacent A-Scan signals are judged as target-containing signals, i k ≤ i ≤ i k+1 . Thus, P target horizontal direction regions are obtained as [i pL i pH , 1 < P < K, 1 ≤ p ≤ P, where i PL is the starting trace number of the p-th target, and i pH is the ending trace number of the p-th target;

[0043] 2) Deletion processing

[0044] For the p-th target after the fusion processing, its horizontal direction width W p = (i pH - i pL ). If W p is less than the threshold value WT, then this target is judged as a false target, and the region corresponding to this target is deleted from the target horizontal direction region, and the final target horizontal direction region is H2.

[0045] Furthermore, the specific steps for segmenting the A-Scan signals in the target horizontal direction region and extracting the time domain, frequency domain, and wavelet domain features of each segment are as follows:

[0046] (1) Segment all the A-Scan signals in the target horizontal direction region, and the length of each segment is L1. Then, the lk-th segment signal in the i-th A-Scan signal x1 i (m) is:

[0047] x2 i,lk (r) = x1 i ((lk - 1)*L1 + r), 1 ≤ lk ≤ KL, 1 ≤ r ≤ L1 (8)

[0048] Wherein, i is the channel number, lk is the segment number, r is the signal point number within the segment, and the number of segments KL=floor(M / L1), where floor(·) means rounding down.

[0049] (2) The process of extracting the time domain, frequency domain and wavelet domain features of each signal segment is as follows:

[0050] 1) Time domain characteristics

[0051] The time domain features include 3 features: mean MEAN i,lk , standard deviation STD i,lk and RMS i,lk , the calculation formulas are as follows:

[0052]

[0053]

[0054]

[0055] Among them, x2 i,lk (r) is the A-Scan signal data of the lkth segment of the i-th channel after segment processing;

[0056] (2) Frequency domain characteristics

[0057] The frequency domain feature is the signal energy spectrum, which is calculated as follows:

[0058]

[0059] Among them, k3 is the frequency point number;

[0060] (3) Wavelet domain characteristics

[0061] The wavelet domain feature is the wavelet packet energy spectrum, which is defined as the normalized wavelet packet node energy:

[0062]

[0063] The wavelet packet node energy is:

[0064]

[0065] in, x2 i,lk (r) is the wavelet packet decomposition coefficient, l is the number of decomposition levels, j is the frequency band, and k4 is the length of the decomposition coefficient.

[0066] Furthermore, the steps of obtaining the target vertical area in the A-Scan signal are as follows:

[0067] (1) Multiple segments of signals containing and excluding hole targets are selected from the A-Scan signals in the horizontal area of ​​the target, and features are extracted to construct a sample set, where each sample is a 19-dimensional feature vector;

[0068] (2) Selecting part of the data from the sample set as the training set, using the training set to train the BP neural network to obtain the network model NET2, and using NET2 to identify the sample set;

[0069] (3) For J segmented signals identified as containing targets By the track number i and the segment number in the track The vertical coordinates of the target area in the i-th A-Scan signal are obtained as

[0070] Furthermore, for the segmented signal identified as containing the target, the corresponding vertical and horizontal regions are marked to obtain the position of the target.

[0071] The beneficial effects brought by the technical solution provided by the present invention are:

[0072] 1. The robust principal component analysis method is used to decompose the ground penetrating radar B-Scan image, which can make full use of the sparse characteristics of the target component and the low-rank characteristics of the clutter component in the image to better separate the target and clutter;

[0073] 2. Selecting the A-Scan signal time domain statistical features, frequency domain energy spectrum and wavelet packet energy spectrum in the wavelet domain as features can make full use of the complementarity between different types of features to better distinguish between targets and non-targets;

[0074] 3. The BP neural network is used to first perform overall recognition of the A-Scan signal, which can quickly identify the target horizontal area; then the A-Scan signal in the target horizontal area is segmented and recognized, which can accurately identify the target position, avoiding the calculation of a large number of non-target areas and improving the recognition efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0076] Figure 1 It is a flow chart of a method for identifying underground cavity targets by ground penetrating radar based on BP neural network in an embodiment of the present invention.

[0077] Figure 2 Schematic diagram of the original B-Scan image of the ground penetrating radar in an embodiment of the present invention.

[0078] Figure 3Schematic diagram of a B-Scan image after clutter is suppressed using a robust principal component analysis method in an embodiment of the present invention.

[0079] Figure 4 Schematic diagram of the target horizontal initial area identified in an embodiment of the present invention.

[0080] Figure 5 It is a schematic diagram of the target horizontal area after fusion processing and deletion processing in an embodiment of the present invention.

[0081] Figure 6 is a schematic diagram of a target position identified in an embodiment of the present invention. DETAILED DESCRIPTION

[0082] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0083] The embodiment of the present invention provides a method for identifying underground cavity targets using a ground penetrating radar based on a BP neural network.

[0084] Please refer to Figure 1 , Figure 1 1 is a flow chart of a method for identifying underground cavity targets by ground penetrating radar based on BP neural network in an embodiment of the present invention, which specifically includes the following steps:

[0085] 1. Use the robust principal component analysis method to process the ground penetrating radar B-Scan image to suppress clutter signals; the specific steps are as follows:

[0086] (1) Obtain the ground penetrating radar B-Scan image X∈R M×N , where N is the number of channels, M is the number of sampling points for each channel, X = [x 1 (m),x 2 (m),…,x N (m)],x i (m) = (x i (1),x i (2),…,x i (M)) T is the i-th A-Scan signal, m represents the sample point number in each A-Scan signal data, 1≤m≤M;

[0087] (2) Each A-Scan signal in the ground penetrating radar B-Scan image X is processed to remove the DC offset. The processing process is as follows:

[0088]

[0089] in, It represents the mean value of each A-Scan signal data in the ground penetrating radar B-Scan image X.

[0090] (3) The robust principal component analysis method is used to process the ground penetrating radar B-scan image after removing the DC offset to obtain a low-rank matrix G and a sparse matrix S. The sparse matrix S is selected as the target echo signal to obtain the B-scan image X1 after clutter suppression.

[0091] 2. Perform feature extraction on each A-Scan signal in the processed B-Scan image, and obtain the following time domain, frequency domain and wavelet domain features:

[0092] (1) Time domain characteristics

[0093] The time domain features include 3 features: mean MEAN i , standard deviation STD i and RMS i , the calculation process is as follows:

[0094]

[0095]

[0096]

[0097] (2) Frequency domain characteristics

[0098] The frequency domain feature is the signal energy spectrum, which is calculated as follows:

[0099]

[0100] Wherein k1 is the frequency point number, and in this embodiment, the energy spectra of the first 12 odd frequency points (k1=1, 3, ..., 23) are selected as the frequency domain features.

[0101] (3) Wavelet domain characteristics

[0102] The wavelet domain feature is the wavelet packet energy spectrum T i , the wavelet packet energy spectrum is defined as the normalized wavelet packet node energy:

[0103]

[0104] Wavelet packet node energy e i,l,j for:

[0105]

[0106] Among them, i is the track number, For x1 i(m)'s wavelet packet decomposition coefficients, l is the number of decomposition layers, j is the frequency band, and k2 is the length of the decomposition coefficients.

[0107] In this embodiment, the decomposition layer number l is selected as 3, and thus the wavelet domain feature dimension is 8.

[0108] Combining the above three major types of features, x1 is obtained i (m)'s feature vector dimension is 23.

[0109] 3. Use a BP neural network to perform overall recognition on the A-Scan signals in the B-Scan image after clutter suppression to obtain the target horizontal direction region; the steps to obtain the target horizontal direction region are as follows:

[0110] (1) Select multiple channels of A-Scan signals with and without hole targets from the B-Scan image after clutter suppression, extract features to construct a sample set, where each sample is a 23-dimensional feature vector, including 3 in the time domain, 12 in the frequency domain, and 8 in the wavelet domain;

[0111] (2) Select part of the data from the sample set as the training set, use the training set to train the BP neural network to obtain the network model NET1, and use NET1 to identify the sample set;

[0112] (3) For the K A-Scan signals identified as containing targets According to their horizontal position i k , obtain the initial target horizontal direction region H1 = {i k , 1 ≤ k ≤ K};

[0113] (4) Perform fusion and deletion processing on the initial target horizontal direction region H1, and the specific process is as follows:

[0114] 1) Fusion processing

[0115] Assume and are two adjacent A-Scan signals containing targets. If their channel difference 1 < |i k - i k+1 | ≤ id, then all the A-Scan signals x1 i (m) between these adjacent A-Scan signals are judged as target-containing signals, i k ≤ i ≤ i k+1 , and thus P target horizontal direction regions are obtained as [i pL i pH , 1 < P < K, 1 ≤ p ≤ P, where i PL is the starting channel number of the p-th target, and i pH is the ending channel number of the p-th target;

[0116] 2) Deletion processing

[0117] For the pth target after fusion processing, its horizontal width W p =(i pH -i pL ), if W p If the target is smaller than the threshold value WT, the target is judged as a false target, and the area corresponding to the target is deleted from the target horizontal direction area, and the final target horizontal direction area is H2.

[0118] 4. Segment the A-Scan signal in the target horizontal area and extract the time domain, frequency domain and wavelet domain features of each segment. The specific steps are as follows:

[0119] (1) Divide all A-Scan signals in the target horizontal area into segments, each segment length is L1, then the i-th A-Scan signal x1 i The lkth segment signal in (m) is:

[0120] x2 i,lk (r) = x1 i ((lk-1)*L1+r), 1≤lk≤KL, 1≤r≤L1 (8)

[0121] Where i is the channel number, lk is the segment number, r is the signal point number in the segment, and the number of segments KL = floor(M / L1), where floor(·) means rounding down.

[0122] (2) The process of extracting the time domain, frequency domain and wavelet domain features of each signal segment is as follows:

[0123] 1) Time domain characteristics

[0124] The time domain features include 3 features: mean MEAN i,lk , standard deviation STD i,lk and RMS i,lk , the calculation formulas are as follows:

[0125]

[0126]

[0127]

[0128] Among them, x2 i,lk (r) is the A-Scan signal data of the lkth segment of the i-th channel after segment processing;

[0129] (2) Frequency domain characteristics

[0130] The frequency domain feature is the signal energy spectrum, which is calculated as follows:

[0131]

[0132] Wherein, k3 is the frequency point number. In this embodiment, the energy spectra of the first 8 frequency points (k3=1, 2, ..., 8) are selected as the frequency domain features;

[0133] (3) Wavelet domain characteristics

[0134] The wavelet domain feature is the wavelet packet energy spectrum T i,lk , the wavelet packet energy spectrum is defined as the normalized wavelet packet node energy:

[0135]

[0136] The wavelet packet node energy is:

[0137]

[0138] in, x2 i,lk (r) is the wavelet packet decomposition coefficient, l is the number of decomposition levels, j is the frequency band, and k4 is the length of the decomposition coefficient.

[0139] In this embodiment, the decomposition level l is selected to be 3, thereby obtaining a wavelet domain feature dimension of 8. Combining the above three types of features, the feature vector dimension of the segmented A-Scan signal is 19.

[0140] 5. According to the time domain, frequency domain and wavelet domain features of each signal obtained in step 4, the BP neural network is used to identify the segmented signals in the A-Scan to obtain the target vertical direction area in the A-Scan signal. The specific steps are as follows:

[0141] (1) Multiple segments of signals with and without hole targets are selected from the A-Scan signals in the horizontal area of ​​the target, and features are extracted to construct a sample set, where each sample is a 19-dimensional feature vector, including 3 in the time domain, 8 in the frequency domain, and 8 in the wavelet domain;

[0142] (2) Selecting part of the data from the sample set as the training set, using the training set to train the BP neural network to obtain the network model NET2, and using NET2 to identify the sample set;

[0143] (3) For J segmented signals identified as containing targets By the track number i and the segment number in the track The vertical coordinates of the target area in the i-th A-Scan signal are obtained as

[0144] 6. According to the horizontal and vertical regions of the target, the segmented signal identified as containing the target is marked in its corresponding vertical and horizontal regions to obtain the position of the target.

[0145] Combination Figure 1 ,by Figure 2 The target recognition method of the embodiment of the present invention is described by taking the original B-Scan image of the ground penetrating radar as an example, and includes the following steps:

[0146] 1. Input the original B-Scan image of the ground penetrating radar, such as Figure 2 As shown, it contains 551 A-Scan signals, and each channel has 512 sampling points;

[0147] 2. Use the robust principal component analysis method to suppress the clutter in the original B-Scan image and obtain the image after clutter suppression, such as Figure 3 As shown;

[0148] 3. For each A-Scan signal, extract the time domain, frequency domain and wavelet domain features and construct a 23-dimensional feature vector;

[0149] 4. The constructed BP neural network model NET1 has 3 layers and a structure of 23-10-2. A sample set is constructed using multiple feature vectors containing and excluding target A-Scan signals to train NET1. The sample set is tested using the trained network to obtain the initial target horizontal area H1, as shown in Figure 4 As shown, the recognition accuracy is 82.6%;

[0150] 5. The identified target horizontal initial area H1 is fused and deleted to obtain the final target horizontal area H2, as shown in Figure 5 As shown, the recognition accuracy is 97.5%;

[0151] 6. Segment the A-Scan signal in the target horizontal area H2, extract the time domain, frequency domain and wavelet domain features of the segmented signal, and construct a 19-dimensional feature vector;

[0152] 7. The constructed BP neural network model NET2 has 3 layers and a structure of 19-10-2. The sample set is constructed using the feature vectors of multiple A-Scan segmented signals containing and excluding targets to train NET2. The sample set is tested using the trained network to obtain the target vertical area V1 in each A-Scan signal.

[0153] 8. For the segmented signals identified as containing targets, mark the corresponding vertical and horizontal areas to obtain the location of the target, such as Figure 6 As shown, the recognition accuracy is 95.5%.

[0154] Compared with the prior art, the present invention has the following beneficial effects:

[0155] 1. The robust principal component analysis method is used to decompose the ground penetrating radar B-Scan image, which can make full use of the sparse characteristics of the target component and the low-rank characteristics of the clutter component in the image to better separate the target and clutter;

[0156] 2. Selecting the A-Scan signal time domain statistical features, frequency domain energy spectrum and wavelet packet energy spectrum in the wavelet domain as features can make full use of the complementarity between different types of features to better distinguish between targets and non-targets;

[0157] 3. The BP neural network is used to first perform overall recognition of the A-Scan signal, which can quickly identify the target horizontal area; then the A-Scan signal in the target horizontal area is segmented and recognized, which can accurately identify the target position, avoiding the calculation of a large number of non-target areas and improving the recognition efficiency.

[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying underground cavity targets using ground penetrating radar based on BP neural network, characterized in that: The following steps are involved: S1: Acquire the ground penetrating radar B-Scan image and process it using the robust principal component analysis method to suppress clutter signals; The specific steps are as follows: (1) Obtain the ground penetrating radar B-Scan image X∈R M×N , where N is the number of A-Scan signals in the B-Scan image, M is the number of sampling points of each A-Scan signal, X = [x1(m), x2(m),…, x N (m)],x i (m) = (x i (1),x i (2),…,x i (M)) T is the i-th A-Scan signal, m represents the sample point number in each A-Scan signal, 1≤m≤M; (2) Each A-Scan signal in the ground penetrating radar B-Scan image X is processed to remove the DC offset. The processing process is as follows: in, It represents the mean value of each A-Scan signal data in the ground penetrating radar B-Scan image X; (3) The robust principal component analysis method is used to process the ground penetrating radar B-Scan image after removing the DC offset to obtain a low-rank matrix G and a sparse matrix S. The sparse matrix S is selected as the target echo signal to obtain the B-Scan image X1 after clutter suppression; S2: Extract features from each A-Scan signal in the processed B-Scan image to obtain features in the time domain, frequency domain, and wavelet domain respectively; S3: Using BP neural network to identify each A-Scan signal in the B-Scan image after clutter suppression, and obtain the target horizontal area; The specific steps are as follows: (1) Multiple A-Scan signals with and without hole targets are selected from the B-Scan image after clutter suppression, and features are extracted to construct a sample set, where each sample is a 23-dimensional feature vector; (2) selecting part of the data from the sample set as a training set, using the training set to train the BP neural network to obtain a network model NET1, and using NET1 to identify the sample set; (3) For K A-Scan signals identified as containing targets According to its horizontal position i k , and obtain the target horizontal initial area H1={i k ,1≤k≤K}; (4) The initial area H1 in the horizontal direction of the target is fused and deleted. The specific process is as follows: 1) Fusion processing Hypothesis and are two adjacent target A-Scan signals. If the trace difference 1 < |i k - i k+1 | ≤ id, then all A-Scan signals x1 i (m) between these adjacent A-Scan signals are judged as target-containing signals, i k ≤ i ≤ i k+1 . Thus, P target horizontal regions are obtained as [i pL i pH , 1 < P < K, 1 ≤ p ≤ P, where i PL is the starting trace number of the p-th target, and i pH is the ending trace number of the p-th target; 2) Deletion processing For the pth target after fusion processing, its horizontal width W p =(i pH -i pL ), if W p If the target is smaller than the threshold value WT, the target is judged as a false target, and the area corresponding to the target is deleted from the target horizontal direction area, and the final target horizontal direction area is H2; S4: segmenting the A-Scan signal in the target horizontal area, and extracting the time domain, frequency domain and wavelet domain features of each segment of the signal; The specific steps are as follows: (1) Divide all A-Scan signals in the target horizontal area into segments, each segment length is L1, then the i-th A-Scan signal x1 i The lkth segment signal in (m) is: x2 i,lk (r)=x1 i ((lk-1)*L1+r),1≤lk≤KL,1≤r≤L1 (2) Where i is the channel number, lk is the segment number, r is the signal point number in the segment, and the number of segments KL = floor(M / L1), where floor(·) means rounding down. (2) The process of extracting the time domain, frequency domain and wavelet domain features of each signal segment is as follows: 1) Time domain characteristics The time domain features include 3 features: mean MEAN i,lk , standard deviation STD i,lk and RMS i,lk , the calculation formulas are as follows: Among them, x2 i,lk (r) is the A-Scan signal data of the lkth segment of the i-th channel after segment processing; (2) Frequency domain characteristics The frequency domain feature is the signal energy spectrum, which is calculated as follows: Where k3 is the frequency point number; (3) Wavelet domain characteristics The wavelet domain feature is the wavelet packet energy spectrum, which is defined as the normalized wavelet packet node energy: The wavelet packet node energy is: in, x2 i,lk (r) is the wavelet packet decomposition coefficient, l is the number of decomposition levels, j is the frequency band, and k4 is the length of the decomposition coefficient; S5: According to the time domain, frequency domain and wavelet domain features of each segment of the signal obtained in step S4, the segmented signals in the A-Scan are identified using a BP neural network to obtain the target vertical direction area in the A-Scan signal; The specific steps are as follows: (1) Multiple segments of signals containing and excluding hole targets are selected from the A-Scan signals in the horizontal area of ​​the target, and features are extracted to construct a sample set, where each sample is a 19-dimensional feature vector; (2) Selecting part of the data from the sample set as the training set, using the training set to train the BP neural network to obtain the network model NET2, and using NET2 to identify the sample set; (3) For J segmented signals identified as containing targets By the track number i and the segment number in the track The vertical coordinates of the target area in the i-th A-Scan signal are obtained as S6: Determine the target position according to the target horizontal direction area and the target vertical direction area.

2. The method for identifying underground cavity targets by ground penetrating radar based on BP neural network as claimed in claim 1, characterized in that: In step S2, the time domain, frequency domain and wavelet domain features obtained are as follows: (1) Time domain characteristics The time domain features include 3 features: mean MEAN i , standard deviation STD i and RMS i , the calculation process is as follows: Among them, x1 i (m) is the i-th A-Scan signal data in the B-Scan image after clutter suppression; (2) Frequency domain characteristics The frequency domain feature is the signal energy spectrum P i (k1), which is calculated as follows: Among them, k1 is the frequency point number; (3) Wavelet domain characteristics The wavelet domain feature is the wavelet packet energy spectrum T i , the wavelet packet energy spectrum is defined as the normalized wavelet packet node energy: Wavelet packet node energy e i,l,j for: Among them, i is the track number, For x1 i (m) is the wavelet packet decomposition coefficient, l is the number of decomposition levels, j is the frequency band, and k2 is the length of the decomposition coefficient.

3. The method for identifying underground cavity targets by ground penetrating radar based on BP neural network as claimed in claim 1, characterized in that: In step S6, the segmented signal identified as containing the target is marked in its corresponding vertical and horizontal regions to obtain the position of the target.

Images