Device and method for accurately acquiring information on anomalies at the working face using multi-frequency ground penetrating radar
By combining multi-frequency ground-penetrating radar (GPR) devices with deep learning, the problem of insufficient accuracy of GPR in complex geological structures has been solved, enabling efficient and accurate identification of underground anomalies and improving detection accuracy and robustness.
Patent Information
- Application Number
- CN202511387412.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-26
- Publication Date
- 2025-12-02
- Estimated Expiration
- 2045-09-26
AI Technical Summary
Existing ground-penetrating radar technology struggles to balance the accuracy of shallow and deep targets when detecting complex geological structures, and its ability to identify anomalies is limited. Traditional fusion methods cannot dynamically adjust weights based on the characteristics of different frequency components, and deep learning models lack robustness and accuracy in detecting underground anomalies.
By employing a multi-frequency ground-penetrating radar device and combining multi-frequency data complementarity with deep learning optimization, and through multi-frequency data acquisition, data alignment, data fusion, model training optimization, and anomaly information identification modules, the system utilizes an improved NSCT transform and an improved YOLOv11 model to dynamically adjust weights and feature extraction, thereby improving the accuracy and efficiency of anomaly detection.
It significantly improves the detection accuracy of anomalies such as cavities, faults, and collapse columns at the working face, enhances robustness and reliability in complex geological backgrounds, and enables efficient and accurate identification of underground anomalies.
Smart Images
Figure CN120871123B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a device for accurately acquiring information on anomalies at a working face, specifically to a device and method for accurately acquiring information on anomalies at a working face using multi-frequency ground penetrating radar. Background Technology
[0002] With the continuous expansion of underground resource mining, such as coal, the importance of working face detection technology in ensuring mining safety and improving efficiency is becoming increasingly prominent. Traditional geological exploration methods (such as drilling and seismic exploration), while providing some geological information, suffer from low detection accuracy, high cost, and poor real-time performance. In recent years, Ground Penetrating Radar (GPR) technology has gradually become an important means of working face detection due to its advantages such as high resolution, non-contact detection, and real-time performance. However, single-frequency GPR often struggles to achieve the same detection accuracy for both shallow and deep targets when detecting complex geological structures, and its ability to identify anomalies is limited. High-frequency GPR data imaging has higher resolution but lower penetration depth, while low-frequency GPR data imaging can provide lower resolution but higher penetration depth. Therefore, multi-frequency GPR technology has emerged, which, by combining radar data from different frequencies, can achieve more comprehensive geological information acquisition across different depth ranges.
[0003] Non-Subsampled Contourlet Transform (NSCT) fusion methods often employ fixed rules or weights when fusing different subbands, failing to adequately consider the varying contributions of different frequency components in ground-penetrating radar (GPR) data to anomaly features. For example, certain subsurface anomalies may exhibit distinct edge features in high-frequency subbands, while displaying overall texture features in low-frequency subbands. Existing fusion methods cannot dynamically adjust weights based on the importance of these features, resulting in fused images that fail to highlight key features and negatively impact anomaly identification.
[0004] In recent years, machine learning technology has made groundbreaking progress in fields such as image recognition and object detection. Especially driven by deep learning, models based on convolutional neural networks (CNNs) have become mainstream. Deep learning, by constructing multi-layered neural networks, can automatically learn complex features in data, thereby achieving efficient pattern recognition and classification. However, traditional object detection models (such as YOLOv3 and YOLOv4) still have certain limitations when handling complex backgrounds and multi-scale targets, especially in specific scenarios such as underground anomaly detection, where the robustness and accuracy of these models still need further improvement. Summary of the Invention
[0005] To address the problems existing in the prior art, the present invention provides a device and method for accurately acquiring information on anomalies on the working face using multi-frequency ground penetrating radar. Through multi-frequency data complementarity and deep learning optimization, the accuracy and efficiency of anomaly detection are significantly improved.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a device for accurately acquiring information on anomalies on a working face using multi-frequency ground penetrating radar, comprising a multi-frequency data acquisition module, a data alignment module, a data fusion module, a multi-frequency data management module, a model training and optimization module, and an anomaly information identification module connected in sequence;
[0007] The multi-frequency data acquisition module is used to calibrate ground-penetrating radar data at different center frequencies, performing operations such as time zero-point correction, gain adjustment, filtering and noise reduction, background elimination, and offset processing to ensure data quality.
[0008] The data alignment module is used to spatially align the preprocessed ground-penetrating radar data to ensure the consistency of radar data of different frequencies in the spatial coordinate system.
[0009] The data fusion module is used to fuse spatially aligned ground-penetrating radar data using improved NSCT to generate multi-frequency ground-penetrating radar fusion profiles, ensuring that the characteristics of shallow and deep anomalies are highlighted.
[0010] The multi-frequency data management module is used to collect multi-frequency ground-penetrating radar fused profile images, label underground anomalies, and create ground-penetrating radar image datasets.
[0011] The model training and optimization module is used to train the improved YOLOv11 model using the labeled ground-penetrating radar image dataset, and to adjust the model parameters using relevant optimization algorithms to improve the model's recognition accuracy and efficiency.
[0012] The anomaly information identification module is used to identify the initial type of underground anomalies and the initial location information of target diseases using an improved YOLOv11 model, and to determine the final type of anomalies and the final location information of target diseases.
[0013] Furthermore, this includes the following steps:
[0014] S10: Use multi-frequency ground-penetrating radar to collect ground-penetrating radar data of different center frequencies of the target;
[0015] S20: Preprocess the raw multi-frequency ground-penetrating radar data. Preprocessing includes time zero-point correction, gain adjustment, filtering and noise reduction, background elimination, and offset processing.
[0016] S30: Spatial alignment of the preprocessed ground-penetrating radar data;
[0017] S40: Perform NSCT data fusion on the ground penetrating radar data processed in step S30 to obtain a multi-frequency ground penetrating radar fusion profile;
[0018] S50: Collect multi-frequency ground-penetrating radar fusion profile images to create a ground-penetrating radar image dataset, which includes a training set, a validation set, and a test set;
[0019] S60: Import the improved YOLOv11 model to identify and label underground anomalies in the ground penetrating radar image dataset.
[0020] Furthermore, in step S10, the target is the working face to be mined, and the ground-penetrating radar data refers to the ground-penetrating radar data extracted from a specified area at different frequencies.
[0021] Furthermore, the specific method of step S20 is as follows:
[0022] S201: Time zero point correction, calibrates the time start point of radar data response at different frequencies, and takes the smallest time zero point as the system time zero point;
[0023] S202: Adjust the gain. Based on zero-point correction, use different exponential gains for different time windows of ground-penetrating radar data of the same frequency for gain processing.
[0024] S203: Filtering and denoising. Based on gain adjustment, bandpass filter denoising technology is used. The radar main frequency is used as the center frequency, and 0.5-1.5 times the center frequency is used as the bandwidth to perform bandpass filtering on ground penetrating radar data collected at the same frequency.
[0025] S204: Background elimination. Based on filtering and noise reduction, the moving average method is used to average the multi-channel radar data into a single data point using a sliding rectangular window.
[0026] S205: Offset processing. Based on background removal, Kirchhoff offset is used to process the time-domain reflection signal of the ground penetrating radar data, remove hyperbolic diffraction effects, and restore the working surface anomaly in the radar image to its actual position.
[0027] Furthermore, the specific method of step S30 is as follows:
[0028] S301: Record the acquisition sequence number and time window of ground penetrating radar data at different antenna center frequencies, perform linear interpolation pre-encryption on sparse data, ensure that the original channel spacing is appropriate, and make the sampling points of the interpolated ground penetrating radar data evenly distributed.
[0029] S302: Perform spatial horizontal alignment to ensure that radar data from different frequencies are in the same location when the acquisition sequence number is the same. The formula is as follows:
[0030] ;
[0031] In the formula: assuming , Indicates the collection sequence number; This represents the radar data coordinates of a characteristic channel. This represents the radar data coordinates after horizontal alignment in the channel space, where ; Represents the floor function; Represents the interpolation function; It represents a positive integer value;
[0032] S303: Perform spatial vertical alignment to ensure that radar data of different frequencies have the same sampling rate and number of sampling points when the time window is the same. The formula is as follows:
[0033] ;
[0034] In the formula: assuming , Indicates a time window; This represents the number of sampling points after vertical alignment of the channel.
[0035] Furthermore, the specific method of step S40 is as follows:
[0036] S401: Perform NSCT transformation on the ground penetrating radar data with different center frequencies after processing in step S30, and decompose it into high-frequency subband and low-frequency subband.
[0037] NSCT consists of NSP and NSDFB. First, NSP decomposes the ground-penetrating radar data into a low-frequency subband and a high-frequency subband, resulting in an NSP. Then, NSDFB is used to further decompose the high-frequency subband. Level direction transformation, to obtain A high-frequency image, this process is an NSCT transform level. The decomposition process;
[0038] As the number of layers increases, the low-frequency subbands will undergo iterative operations on the NSP. If the image has passed through layers... The NSCT transform will yield a low-frequency image and 1 high-frequency image, of which express exist The number of directions in the order decomposition;
[0039] S402: The low-frequency subband is processed using the regional energy fusion rule, and the high-frequency subband is processed using the least squares fusion rule;
[0040] The specific rules for regional energy integration include:
[0041] For the low-frequency subband, a Gaussian weighted window is selected to calculate the local energy:
[0042] ;
[0043] In the formula, This represents the coefficient matrices of the two low-frequency subbands after NSCT decomposition; Indicates the current pixel coordinate position; Indicates A 5×5 neighborhood window centered at coordinates, where ; Let be a two-dimensional Gaussian kernel function, where ; Indicates the first Low-frequency subband in coordinate The local energy value at that location;
[0044] The least squares fusion rules specifically include:
[0045] extract The high-frequency subband data within a 5×5 window around the coordinate position are used to establish a linear relationship between the high-frequency subbands of the two images within a local window, thus establishing a linear mapping relationship:
[0046] ;
[0047] In the formula, This represents 5×5 window data from two high-frequency subbands; Indicates the bias term; Indicates the slope coefficient; Represents the Gaussian noise term;
[0048] Least squares objective function with dynamic regularization:
[0049] ;
[0050] In the formula, Denotes the estimated regression coefficient vector, where ; This indicates the search for the value that minimizes the objective function. value; This represents the observation vector, which expands all high-frequency subband values within the window into a vector by columns. This represents the design matrix, used to model linear regression relationships; Represents the square of the L2 norm; Represents the adaptive regularization coefficient, which dynamically balances fitting error and parameter stability, where Indicates the signal-to-noise ratio. Represents the computation matrix The condition number; S403: Perform NSCT inverse transform on the processed subband to obtain the multi-frequency ground-penetrating radar fusion profile;
[0051] The inverse NSCT transform restores and reconstructs the decomposed filter. After performing the inverse NSCT transform on the processed high-frequency and low-frequency subbands, the fused multi-frequency ground-penetrating radar profile is obtained.
[0052] Furthermore, the specific method of step S50 is as follows:
[0053] S501: Integrate the ground-penetrating radar images processed in step S40, label them with LabelImg, and use rectangular boxes to label the hyperbolic subsurface anomalies to obtain several labeled ground-penetrating radar images.
[0054] S502: To ensure a reasonable distribution of training, validation, and testing data, the labeled ground-penetrating radar images are divided into training, validation, and testing sets in a 6:2:2 ratio.
[0055] Furthermore, the improved YOLOv11 model in step S60 includes three parts: Backbone, Neck, and DyHead.
[0056] The backbone consists of five Convs, three C3k2s, a GST, an SPPF, and a C2PSA. The first Conv is connected to the second Conv, and the second to fifth Convs are each connected by a C3k2. The GST, SPPF, and C2PSA are connected sequentially, and the fifth Conv is connected to the GST.
[0057] The Neck consists of two Upsamples, four EMA_C3k2s, four Concats, and two Convs. The first Upsample, the first Concat, the first EMA_C3k2, the second Upsample, the second Concat, and the second EMA_C3k2 are connected in sequence. The first Conv, the third Concat, the third EMA_C3k2, the second Conv, the fourth Concat, and the fourth EMA_C3k2 are connected in sequence. The first EMA_C3k2 is connected to the third Concat, and the second EMA_C3k2 is connected to the first Conv.
[0058] DyHead includes three Dytects;
[0059] The second and third C3k2 in the Backbone are connected to the second and first Concat in the Neck, respectively. Similarly, the C2PSA in the Backbone is connected to the first Upsample and the fourth Concat in the Neck.
[0060] The second EMA_C3k2, the third EMA_C3k2, and the fourth EMA_C3k2 in the Neck are connected to the first Dytect, the second Dytect, and the third Dytect in the DyHead, respectively.
[0061] Furthermore, the specific method of step S60 is as follows:
[0062] S601: Ground Penetrating Radar Image Preprocessing; After inputting the ground penetrating radar image into the model, the ground penetrating radar images of different sizes are adjusted to a uniform size of 640×640, while maintaining the original aspect ratio to prevent missing or distorted images from losing their original information.
[0063] S602: Ground Penetrating Radar Image Feature Extraction; The preprocessed image is input into Backbone, and GST is used for further feature extraction. The formula is as follows: ;
[0064] In the formula, The query matrix represents the query vector containing the input features; The key matrix represents the key vectors of the input features; The value matrix represents the value vector of the input features. The dimension of the key is used to scale the dot product and prevent gradient vanishing. This represents the activation function used to calculate attention weights;
[0065] S603: Ground Penetrating Radar Image Feature Fusion; In the Neck section, EMA_C3k2 is used to fuse features by grouping channels and capturing global and local features using global average pooling and convolution operations. The global average pooling formula is as follows: ;
[0066] In the formula, This represents the input feature map, which is the output of the previous layer. Indicates the height of the feature map; Indicates the width of the feature map; Indicates feature map in The value at the coordinates;
[0067] Perform a convolution operation to generate channel weights, as shown in the following formula: ;
[0068] In the formula, This represents the channel weights, used to weight the input feature maps. This represents the convolution kernel, used for feature transformation; Indicates the offset, used to adjust the output offset; This indicates a convolution operation used for feature transformation.
[0069] Channel weights are applied to the input feature map to enhance features relevant to subsurface anomalies and suppress irrelevant features. These features are then fused with the high-resolution feature map, as shown in the following formula: ;
[0070] In the formula, This represents the weighted feature map, and the feature map after processing by the attention mechanism. This represents the activation function, used to introduce nonlinearity; Indicates the final feature of the previous layer; This indicates a feature map concatenation operation, used to merge feature maps of different scales.
[0071] S604: Target recognition; The fused feature map is input into the Head part, and the formula is as follows: ;
[0072] In the formula, Indicates the input feature map; Indicates the horizontal dimension; Indicates spatial dimension; Indicates the channel dimension; Represents matrix multiplication;
[0073] Three DyHead dynamic detection heads were used to process feature maps with resolutions of 20×20, 40×40, and 80×80 in the Neck network to perform target labeling and category prediction for underground anomalies.
[0074] Compared with the prior art, the present invention has the following beneficial effects:
[0075] This invention employs multi-frequency ground-penetrating radar data fusion technology based on NSCT transform, ultimately obtaining a multi-frequency ground-penetrating radar fusion profile. This overcomes the limitations of previous single-frequency ground-penetrating radar detection and significantly improves the detection accuracy of anomalies such as cavities, faults, and collapse columns at the working face.
[0076] This invention incorporates Swing Transformer, EMA, and DyHead into the YOLOv11 model, demonstrating excellent performance in anomaly detection under complex geological backgrounds. The model can efficiently and accurately identify subsurface anomalies, with significant advantages in multi-scale target detection and complex background noise suppression, thereby improving the robustness and reliability of the detection.
[0077] This invention provides a device for accurately acquiring information on anomalies at working faces using multi-frequency ground-penetrating radar. This device integrates multiple functional modules, including multi-frequency data acquisition, data alignment, data fusion, data management, model training and optimization, and anomaly information identification. It realizes the entire process from data acquisition to anomaly identification, providing a complete solution for geological exploration in complex underground environments such as coal mines. It has good application prospects and promotional value. Attached Figure Description
[0078] Figure 1 This is a schematic diagram of the overall process flow of the method of the present invention;
[0079] Figure 2 This is a schematic diagram of the multi-frequency ground-penetrating radar data fusion process based on NSCT transform according to the present invention;
[0080] Figure 3 This invention provides an improved YOLOv11 model network structure diagram;
[0081] Figure 4 This is a diagram of the gprMax model in an embodiment of the present invention;
[0082] Figure 5 for Figure 4 The model was forward-modeled using gprMax software, and the 900MHz center frequency Bscan image is shown.
[0083] Figure 6 for Figure 4 The model was simulated using gprMax software, and the 2.6 GHz center frequency Bscan image is shown.
[0084] Figure 7 For the Figure 5 and Figure 6 Fusion profile of multi-frequency ground-penetrating radar;
[0085] Figure 8 This is a schematic diagram of the device structure of the present invention. Detailed Implementation
[0086] The invention will now be further described with reference to the accompanying drawings.
[0087] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0088] like Figure 8As shown, the present invention provides a device for accurately acquiring information on anomalies on a working face using a multi-frequency ground penetrating radar, comprising a multi-frequency data acquisition module, a data alignment module, a data fusion module, a multi-frequency data management module, a model training and optimization module, and an anomaly information identification module connected in sequence.
[0089] The multi-frequency data acquisition module is used to calibrate ground-penetrating radar data at different center frequencies, performing operations such as time zero-point correction, gain adjustment, filtering and noise reduction, background elimination and offset processing to ensure data quality.
[0090] The data alignment module is used to spatially align the preprocessed ground-penetrating radar data to ensure the consistency of radar data of different frequencies in the spatial coordinate system.
[0091] The data fusion module is used to fuse spatially aligned ground-penetrating radar data using improved NSCT to generate multi-frequency ground-penetrating radar fusion profiles, ensuring that the characteristics of shallow and deep anomalies are highlighted.
[0092] The multi-frequency data management module is used to collect multi-frequency ground-penetrating radar fused profile images, label underground anomalies, and create ground-penetrating radar image datasets.
[0093] The model training and optimization module is used to train the improved YOLOv11 model using the labeled ground-penetrating radar image dataset, and to adjust the model parameters using relevant optimization algorithms to improve the model's recognition accuracy and efficiency.
[0094] The anomaly information identification module is used to identify the initial type of underground anomalies and the initial location information of target diseases using an improved YOLOv11 model, and to determine the final type of anomalies and the final location information of target diseases.
[0095] like Figure 1 As shown, the method for accurately acquiring information about anomalies at the working face using multi-frequency ground-penetrating radar includes the following steps:
[0096] S10: Use multi-frequency ground-penetrating radar to collect ground-penetrating radar data at different center frequencies from the working face to be mined;
[0097] Ground penetrating radar data refers to ground penetrating radar data extracted from a specified area at different frequencies.
[0098] S20: Preprocess the raw multi-frequency ground-penetrating radar data. Preprocessing includes time zero-point correction, gain adjustment, filtering and noise reduction, background elimination, and offset processing.
[0099] S201: Time zero point correction, calibrates the time start point of radar data response at different frequencies, and takes the smallest time zero point as the system time zero point;
[0100] S202: Adjust the gain. Based on zero-point correction, use different exponential gains for different time windows of ground-penetrating radar data of the same frequency for gain processing.
[0101] S203: Filtering and denoising. Based on gain adjustment, bandpass filter denoising technology is used. The radar main frequency is used as the center frequency, and 0.5-1.5 times the center frequency is used as the bandwidth to perform bandpass filtering on ground penetrating radar data collected at the same frequency.
[0102] S204: Background elimination. Based on filtering and noise reduction, the moving average method is used to average the multi-channel radar data into a single data point using a sliding rectangular window, thereby achieving the purpose of background elimination.
[0103] S205: Offset processing. Based on background removal, Kirchhoff offset is used to process the time-domain reflection signal of the ground penetrating radar data, remove hyperbolic diffraction effects, and restore the working surface anomaly in the radar image to its actual position.
[0104] S30: Spatial alignment of preprocessed ground-penetrating radar data.
[0105] S301: Record the acquisition sequence number and time window of ground penetrating radar data at different antenna center frequencies, perform linear interpolation pre-encryption on sparse data, ensure that the original channel spacing is appropriate, and make the sampling points of the interpolated ground penetrating radar data evenly distributed.
[0106] S302: Perform spatial horizontal alignment to ensure that radar data from different frequencies are in the same location when the acquisition sequence number is the same. The formula is as follows:
[0107] In the formula: assuming , Indicates the collection sequence number; This represents the radar data coordinates of a characteristic channel. This represents the radar data coordinates after horizontal alignment in the channel space, where ;
[0108] Represents the floor function; Represents the interpolation function; It represents a positive integer value;
[0109] S303: Perform spatial vertical alignment to ensure that radar data of different frequencies have the same sampling rate and number of sampling points when the time window is the same. The formula is as follows:
[0110] ;
[0111] In the formula: assuming , Indicates a time window; This represents the number of sampling points after vertical alignment of the channel.
[0112] S40: As Figure 2 As shown, the ground penetrating radar data processed in step S30 is subjected to NSCT, i.e., non-subsampled contour wave transform data fusion to obtain a multi-frequency ground penetrating radar fusion profile.
[0113] S401: Perform NSCT transformation on the ground penetrating radar data with different center frequencies after processing in step S30, and decompose it into high-frequency subband and low-frequency subband.
[0114] The NSCT consists of an NSP (Non-Subsampled Tower Filter) and an NSDFB (Non-Subsampled Directional Filter). First, the NSP decomposes the ground-penetrating radar data into a multi-scale and multi-resolution subband, resulting in a low-frequency subband and a high-frequency subband. Then, the NSDFB is used to further refine the high-frequency subband. Level direction transformation, to obtain A high-frequency image, this process is an NSCT transform level. The decomposition process;
[0115] As the number of layers increases, the low-frequency subbands will undergo iterative operations on the NSP. If the image has passed through layers... The NSCT transform will yield a low-frequency image and 1 high-frequency image, of which express exist Number of directions in the decomposition; in this example, NSCT decomposition yields 1 low-frequency subband and 8 high-frequency subbands in 3 directions, i.e., after 3-level decomposition;
[0116] S402: The low-frequency subband is processed using the regional energy fusion rule, and the high-frequency subband is processed using the least squares fusion rule;
[0117] The specific rules for regional energy integration include:
[0118] For the low-frequency subband, a Gaussian weighted window is selected to calculate the local energy:
[0119] ;
[0120] In the formula, This represents the coefficient matrices of the two low-frequency subbands after NSCT decomposition; Indicates the current pixel coordinate position; Indicates A 5×5 neighborhood window centered at coordinates, where ; Let be a two-dimensional Gaussian kernel function, where ; Indicates the first Low-frequency subband in coordinate The local energy value at that location;
[0121] The least squares fusion rules specifically include:
[0122] extract The high-frequency subband data within a 5×5 window around the coordinate position are used to establish a linear relationship between the high-frequency subbands of two images within a local window. The purpose is to capture the spatial correlation of high-frequency details and establish a linear mapping relationship.
[0123] ;
[0124] In the formula, This represents 5×5 window data from two high-frequency subbands; Indicates the bias term; Indicates the slope coefficient; Represents the Gaussian noise term;
[0125] The least-squares objective function with dynamic regularization is introduced to balance noise suppression and detail preservation, and to avoid overfitting.
[0126] ;
[0127] In the formula, Denotes the estimated regression coefficient vector, where ; This indicates the search for the value that minimizes the objective function. value; This represents the observation vector, which expands all high-frequency subband values within the window into a vector by columns. This represents the design matrix, used to model linear regression relationships; Represents the square of the L2 norm; Represents the adaptive regularization coefficient, which dynamically balances fitting error and parameter stability, where Indicates the signal-to-noise ratio. Represents the computation matrix condition number;
[0128] S403: Perform NSCT inverse transform on the processed subband to obtain a multi-frequency ground-penetrating radar fusion profile;
[0129] The inverse NSCT transform restores and reconstructs the decomposed filter. After performing the inverse NSCT transform on the processed high-frequency and low-frequency subbands, the fused multi-frequency ground-penetrating radar profile is obtained.
[0130] S50: Collect multi-frequency ground-penetrating radar fusion profile images to create a ground-penetrating radar image dataset, which includes a training set, a validation set, and a test set.
[0131] S501: Integrate the ground-penetrating radar images processed in step S40, label them with LabelImg, and use rectangular boxes to label hyperbolic underground anomalies (such as cavities) to obtain several labeled ground-penetrating radar images.
[0132] S502: To ensure a reasonable distribution of training, validation, and testing data, the labeled ground-penetrating radar images are divided into training, validation, and testing sets in a 6:2:2 ratio.
[0133] S60: Import the improved YOLOv11 model to identify and label underground anomalies in the ground penetrating radar image dataset.
[0134] like Figure 3 As shown, the improved YOLOv11 model consists of three parts: Backbone, Neck, and DyHead.
[0135] The backbone consists of five Conv (convolutional modules), three C3k2 (feature extraction modules), a GST (multi-scale extraction module), an SPPF (spatial pyramid pooling module), and a C2PSA (cross-stage attention module). The first Conv is connected to the second Conv, and the second to fifth Convs are each connected by a C3k2 module. The GST, SPPF, and C2PSA modules are connected sequentially, and the fifth Conv is connected to the GST.
[0136] The Neck consists of two Upsample modules, four EMA_C3k2 modules, four Concat modules, and two Conv modules. The first Upsample, the first Concat, the first EMA_C3k2, the second Upsample, the second Concat, and the second EMA_C3k2 are connected sequentially. The first Conv, the third Concat, the third EMA_C3k2, the second Conv, the fourth Concat, and the fourth EMA_C3k2 are connected sequentially. The first EMA_C3k2 is connected to the third Concat module, and the second EMA_C3k2 is connected to the first Conv module.
[0137] DyHead includes three Dytect (dynamic detection head modules);
[0138] The second and third C3k2 in the Backbone are connected to the second and first Concat in the Neck, respectively. Similarly, the C2PSA in the Backbone is connected to the first Upsample and the fourth Concat in the Neck.
[0139] The second EMA_C3k2, the third EMA_C3k2, and the fourth EMA_C3k2 in the Neck are connected to the first Dytect, the second Dytect, and the third Dytect in the DyHead, respectively.
[0140] In the backbone, a Swin Transformer replaces part of the original Conv function, and combined with convolutional operations, a GST is constructed to replace part of the C3k2 function, aiming to enhance the model's ability to extract multi-scale features. Compared to previous models that used C3K2 to handle feature extraction at different stages of the backbone, the YOLOv11 model uses C3K2 to segment feature maps and apply a series of smaller kernel convolutions (3x3) to optimize the information flow in the network, which is faster and less computationally expensive than larger kernel convolutions. The Swin Transformer, through its windowed self-attention mechanism and shifted window strategy, can effectively capture long-range dependencies and enhance multi-scale feature extraction capabilities. When processing ground-penetrating radar images, this improvement can better capture the features of underground anomalies, especially performing well in complex backgrounds and multi-scale target detection.
[0141] The Neck module incorporates an Efficient Multi-scale Attention (EMA) mechanism. The EMA_C3k2 module combines the EMA and C3k2 modules. The EMA_C3k2 module captures global and local features by grouping channels and utilizing global average pooling and convolution operations, achieving channel weight recalibration and pixel-level relationship capture with lower computational cost. This improvement enhances feature fusion capabilities, enabling the model to better focus on key features of subsurface anomalies and reduce background noise interference.
[0142] DyHead replaces the original Head (detection head network). By integrating scale-aware, spatial-aware, and task-aware attention mechanisms, DyHead can introduce attention mechanisms into each dimension of the feature tensor, unifying and enhancing feature representation capabilities. This improvement can enhance the model's detection accuracy for underground anomalies of different scales and shapes while maintaining high computational efficiency.
[0143] S601: Ground penetrating radar image preprocessing;
[0144] After inputting the ground-penetrating radar images into the model, the ground-penetrating radar images of different sizes are adjusted to a uniform size of 640×640, while maintaining the original aspect ratio to prevent missing or distorted images from losing their original information.
[0145] S602: Ground penetrating radar image feature extraction;
[0146] The preprocessed image is input into the improved Backbone, and GST is used for further feature extraction, as shown in the following formula:
[0147] ;
[0148] In the formula, The query matrix represents the query vector containing the input features; The key matrix represents the key vectors of the input features; The value matrix represents the value vector of the input features. The dimension of the key is used to scale the dot product and prevent gradient vanishing. This represents the activation function used to calculate attention weights;
[0149] S603: Ground penetrating radar image feature fusion;
[0150] In the Neck section, EMA_C3k2 is used to fuse features. This is achieved by grouping channels and capturing global and local features using global average pooling and convolution operations. The formula for global average pooling is as follows:
[0151] ;
[0152] In the formula, This represents the input feature map, which is usually the output of the previous layer. Indicates the height of the feature map; Indicates the width of the feature map; Indicates feature map in The value at the coordinates;
[0153] Perform a convolution operation to generate channel weights, as shown in the following formula:
[0154] ;
[0155] In the formula, This represents the channel weights, used to weight the input feature maps. This represents the convolution kernel, used for feature transformation; Indicates the offset, used to adjust the output offset; This indicates a convolution operation used for feature transformation.
[0156] Channel weights are applied to the input feature map to enhance features related to subsurface anomalies (such as edges, textures, and hyperbolic features) and suppress irrelevant features (background noise). This is then fused with a high-resolution (160×160 - 320×320) feature map, as shown in the following formula:
[0157] In the formula, This represents the weighted feature map, and the feature map after processing by the attention mechanism. This represents the activation function, used to introduce nonlinearity; Indicates the final feature of the previous layer; This indicates a feature map concatenation operation, used to merge feature maps of different scales.
[0158] S604: Target recognition;
[0159] The fused feature map is then input into the improved Head part, as shown in the following formula:
[0160] In the formula, Indicates the input feature map; Indicates the horizontal dimension; Indicates spatial dimension; Indicates the channel dimension; Represents matrix multiplication;
[0161] Three DyHead dynamic detection heads were used to process feature maps with resolutions of 20×20, 40×40, and 80×80 in the Neck network to perform target labeling and category prediction for underground anomalies.
[0162] Example:
[0163] Forward modeling was performed using gprMax software, such as... Figure 4 The image shown is a gprMax model. Forward modeling was performed using ground-penetrating radar with a center frequency of 900 MHz and 2.6 GHz, generating several Bscan images of underground anomalies. One of the models is shown here for demonstration: This model is 3m long and 2m deep. In the Y-axis direction, there is a 0.3m air layer at 1.7-2m and a 1.7m coal seam at 0-1.7m. In the coal seam, there are three cavities with a radius of 0.02m and a depth of 0.3m at a Y-axis coordinate of 1.4m. In addition, there are three cavities with progressively increasing depths in the deeper part of the coal seam: the radii of the cavities are 0.1m, 0.15m and 0.2m, and the coordinates of the cavities from shallowest to deepest are (1.6, 1.1, 0.0), (2.0, 0.85, 0.0) and (2.5, 0.4, 0.0).
[0164] The Bscan image generated by gprMax software is preprocessed, including time zero-point correction, gain adjustment, filtering and denoising, background removal, and offset processing.
[0165] Spatial alignment is performed on the preprocessed Bscan image;
[0166] The Bscan images processed as described above are fused using NSCT to obtain a multi-frequency ground-penetrating radar fused profile. Figures 5-7The horizontal axis represents the track number, and the vertical axis represents the time window. Figure 7 contrast Figure 5 and Figure 6 It can be seen that the fusion results based on the NSCT fusion algorithm can clearly identify shallow hyperbolas and highly identify deep hyperbolas simultaneously. This means that the method of this invention can clearly identify shallow cavities and can highly identify deep cavities.
[0167] To further verify the effectiveness of the multi-frequency ground-penetrating radar fusion profile, three parameters—information entropy, spatial gradient, and spatial frequency—will be compared. The results are shown in Table 1.
[0168] Table 1
[0169]
[0170] As shown in Table 1, all three parameters are improved in the fused data, especially the information entropy. This indicates that the fused profile contains more information than any single-frequency Bscan image. The above comprehensive evidence proves that the present invention integrates the advantages of low-frequency radar deep cavity information and high-frequency radar shallow cavity information accuracy.
[0171] The ground-penetrating radar images after the above steps are integrated and labeled using the LabelImg tool. Hyperbolic shapes of underground anomalies (such as cavities) are labeled with rectangular boxes to obtain several labeled ground-penetrating radar images. A ground-penetrating radar image dataset is created, which includes a training set, a validation set, and a test set, which are divided into the training set, validation set, and test set in a ratio of 6:2:2.
[0172] Random direction Figure 3 The improved YOLOv11 model shown imports ground-penetrating radar images containing different types of underground anomalies to identify and mark the anomalies.
[0173] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
[0174] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any minor modifications, equivalent substitutions, and improvements made to the above embodiments based on the technical essence of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for accurately acquiring information on anomalies at a working face using multi-frequency ground-penetrating radar, implemented using a device for accurately acquiring information on anomalies at a working face using multi-frequency ground-penetrating radar, the device comprising a multi-frequency data acquisition module, a data alignment module, a data fusion module, a multi-frequency data management module, a model training and optimization module, and an anomaly information identification module connected in sequence; The multi-frequency data acquisition module is used to calibrate ground-penetrating radar data at different center frequencies, performing operations such as time zero-point correction, gain adjustment, filtering and noise reduction, background elimination, and offset processing to ensure data quality. The data alignment module is used to spatially align the preprocessed ground-penetrating radar data to ensure the consistency of radar data of different frequencies in the spatial coordinate system. The data fusion module is used to fuse spatially aligned ground-penetrating radar data using improved NSCT to generate multi-frequency ground-penetrating radar fusion profiles, ensuring that the characteristics of shallow and deep anomalies are highlighted. The multi-frequency data management module is used to collect multi-frequency ground-penetrating radar fused profile images, label underground anomalies, and create ground-penetrating radar image datasets. The model training and optimization module is used to train the improved YOLOv11 model using the labeled ground-penetrating radar image dataset, and to adjust the model parameters using relevant optimization algorithms to improve the model's recognition accuracy and efficiency. An anomaly information identification module, used to identify the initial type of underground anomalies and the initial location information of target diseases using an improved YOLOv11 model, and to determine the final type of anomalies and the final location information of target diseases, is characterized by including the following steps: S10: Use multi-frequency ground-penetrating radar to collect ground-penetrating radar data of different center frequencies of the target; S20: Preprocess the raw multi-frequency ground-penetrating radar data. Preprocessing includes time zero-point correction, gain adjustment, filtering and noise reduction, background elimination, and offset processing. S30: Spatial alignment of the preprocessed ground-penetrating radar data; S40: Perform NSCT data fusion on the ground penetrating radar data processed in step S30 to obtain a multi-frequency ground penetrating radar fusion profile; The specific method for step S40 is as follows: S401: Perform NSCT transformation on the ground penetrating radar data with different center frequencies after processing in step S30, and decompose it into high-frequency subband and low-frequency subband. NSCT consists of NSP and NSDFB. First, NSP decomposes the ground-penetrating radar data into a low-frequency subband and a high-frequency subband, resulting in an NSP. Then, NSDFB is used to further decompose the high-frequency subband. Level direction transformation, to obtain A high-frequency image, this process is an NSCT transform level. The decomposition process; As the number of layers increases, the low-frequency subbands will undergo iterative operations on the NSP. If the image has passed through layers... The NSCT transform will yield a low-frequency image and 1 high-frequency image, of which express exist The number of directions on the order decomposition; S402: The low-frequency subband is processed using the regional energy fusion rule, and the high-frequency subband is processed using the least squares fusion rule; The specific rules for regional energy integration include: For the low-frequency subband, a Gaussian weighted window is selected to calculate the local energy: ; In the formula, This represents the coefficient matrices of the two low-frequency subbands after NSCT decomposition; Indicates the current pixel coordinate position; Indicated by A 5×5 neighborhood window centered at coordinates, where ; Let be a two-dimensional Gaussian kernel function, where ; Indicates the first Low-frequency subband in coordinate The local energy value at that location; The least squares fusion rules specifically include: extract The high-frequency subband data within a 5×5 window around the coordinate position are used to establish a linear relationship between the high-frequency subbands of the two images within a local window, thus establishing a linear mapping relationship: ; In the formula, This represents 5×5 window data from two high-frequency subbands; Indicates the bias term; Indicates the slope coefficient; Represents the Gaussian noise term; Least squares objective function with dynamic regularization: ； In the formula, Denotes the estimated regression coefficient vector, where ; This indicates the search for the value that minimizes the objective function. value; This represents the observation vector, which expands all high-frequency subband values within the window into a vector by columns. This represents the design matrix, used to model linear regression relationships; Represents the square of the L2 norm; Represents the adaptive regularization coefficient, which dynamically balances fitting error and parameter stability, where Indicates the signal-to-noise ratio. Represents the computation matrix The condition number; S403: Perform NSCT inverse transform on the processed subband to obtain the multi-frequency ground-penetrating radar fusion profile; The inverse NSCT transform is used to restore and reconstruct the decomposed filter. The high-frequency and low-frequency subbands are then subjected to the inverse NSCT transform to obtain the fused multi-frequency ground-penetrating radar profile. S50: Collect multi-frequency ground-penetrating radar fusion profile images to create a ground-penetrating radar image dataset, which includes a training set, a validation set, and a test set; S60: Import the improved YOLOv11 model to identify and label underground anomalies in the ground penetrating radar image dataset.
2. The method for accurately acquiring information on anomalies at a working face using multi-frequency ground-penetrating radar according to claim 1, characterized in that, In step S10, the target is the working face to be mined, and the ground-penetrating radar data refers to the ground-penetrating radar data extracted from a specified area at different frequencies.
3. The method for accurately acquiring information on anomalies at a working face using multi-frequency ground-penetrating radar according to claim 1, characterized in that, The specific method for step S20 is as follows: S201: Time zero point correction, calibrates the time start point of radar data response at different frequencies, and takes the smallest time zero point as the system time zero point; S202: Adjust the gain. Based on zero-point correction, use different exponential gains for different time windows of ground-penetrating radar data of the same frequency for gain processing. S203: Filtering and denoising. Based on gain adjustment, bandpass filter denoising technology is used. The radar main frequency is used as the center frequency, and 0.5-1.5 times the center frequency is used as the bandwidth to perform bandpass filtering on ground penetrating radar data collected at the same frequency. S204: Background elimination. Based on filtering and noise reduction, the moving average method is used to average the multi-channel radar data into a single data point using a sliding rectangular window. S205: Offset processing. Based on background removal, Kirchhoff offset is used to process the time-domain reflection signal of the ground penetrating radar data, remove hyperbolic diffraction effects, and restore the working surface anomaly in the radar image to its actual position.
4. The method for accurately acquiring information on anomalies at a working face using multi-frequency ground-penetrating radar according to claim 1, characterized in that, The specific method for step S30 is as follows: S301: Record the acquisition sequence number and time window of ground penetrating radar data at different antenna center frequencies, perform linear interpolation pre-encryption on sparse data, ensure that the original channel spacing is appropriate, and make the sampling points of the interpolated ground penetrating radar data evenly distributed. S302: Perform spatial horizontal alignment to ensure that radar data from different frequencies are in the same location when the acquisition sequence number is the same. The formula is as follows: ； In the formula: assuming , Indicates the collection sequence number; This represents the radar data coordinates of a characteristic channel. This represents the radar data coordinates after horizontal alignment in this channel space. in ; Represents the floor function; Represents the interpolation function; It represents a positive integer value; S303: Perform spatial vertical alignment to ensure that radar data of different frequencies have the same sampling rate and number of sampling points when the time window is the same. The formula is as follows: ； In the formula: assuming , Indicates a time window; This represents the number of sampling points after vertical alignment of the channel.
5. The method for accurately acquiring information on anomalies at a working face using multi-frequency ground-penetrating radar according to claim 1, characterized in that, The specific method for step S50 is as follows: S501: Integrate the ground-penetrating radar images processed in step S40, label them with LabelImg, and use rectangular boxes to label the hyperbolic subsurface anomalies to obtain several labeled ground-penetrating radar images. S502: To ensure a reasonable distribution of training, validation, and testing data, the labeled ground-penetrating radar images are divided into training, validation, and testing sets in a 6:2:2 ratio.
6. The method for accurately acquiring information on anomalies at a working face using multi-frequency ground-penetrating radar according to claim 1, characterized in that, The improved YOLOv11 model in step S60 includes three parts: Backbone, Neck, and DyHead. The backbone consists of five Convs, three C3k2s, a GST, an SPPF, and a C2PSA. The first Conv is connected to the second Conv, and the second to fifth Convs are each connected by a C3k2. The GST, SPPF, and C2PSA are connected sequentially, and the fifth Conv is connected to the GST. The Neck consists of two Upsamples, four EMA_C3k2s, four Concats, and two Convs. The first Upsample, the first Concat, the first EMA_C3k2, the second Upsample, the second Concat, and the second EMA_C3k2 are connected in sequence. The first Conv, the third Concat, the third EMA_C3k2, the second Conv, the fourth Concat, and the fourth EMA_C3k2 are connected in sequence. The first EMA_C3k2 is connected to the third Concat, and the second EMA_C3k2 is connected to the first Conv. DyHead includes three Dytects; The second and third C3k2 in the Backbone are connected to the second and first Concat in the Neck, respectively. The C2PSA in the Backbone is connected to the first Upsample and the fourth Concat in the Neck, respectively. The second, third, and fourth EMA_C3k2 in the Neck are connected to the first, second, and third Dytect in the DyHead, respectively.
7. The method for accurately acquiring information on anomalies at a working face using multi-frequency ground-penetrating radar according to claim 6, characterized in that, The specific method for step S60 is as follows: S601: Ground penetrating radar image preprocessing; After inputting the ground-penetrating radar images into the model, the ground-penetrating radar images of different sizes are adjusted to a uniform size of 640×640, while maintaining the original aspect ratio to prevent missing or distorted images from losing their original information. S602: Ground penetrating radar image feature extraction; The preprocessed image is input into Backbone, and GST is used for further feature extraction. The formula is as follows: ； In the formula, The query matrix represents the query vector containing the input features; The key matrix represents the key vectors of the input features; The value matrix represents the value vector of the input features. The dimension of the key is used to scale the dot product and prevent gradient vanishing. This represents the activation function used to calculate attention weights; S603: Ground penetrating radar image feature fusion; In the Neck section, EMA_C3k2 is used to fuse features. This is achieved by grouping channels and capturing global and local features using global average pooling and convolution operations. The global average pooling formula is as follows: ； In the formula, This represents the input feature map, which is the output of the previous layer. Indicates the height of the feature map; Indicates the width of the feature map; Indicates feature map in The value at the coordinates; Perform a convolution operation to generate channel weights, as shown in the following formula: ； In the formula, This represents the channel weights, used to weight the input feature maps. This represents the convolution kernel, used for feature transformation; Indicates the offset, used to adjust the output offset; This indicates a convolution operation used for feature transformation. Channel weights are applied to the input feature map to enhance features relevant to subsurface anomalies and suppress irrelevant features. These features are then fused with the high-resolution feature map, as shown in the following formula: ； In the formula, This represents the weighted feature map, and the feature map after processing by the attention mechanism. This represents the activation function, used to introduce nonlinearity; Indicates the final feature of the previous layer; This indicates a feature map concatenation operation, used to merge feature maps of different scales. S604: Target recognition; The fused feature map is input into the Head part, as shown in the following formula: ； In the formula, Indicates the input feature map; Indicates the horizontal dimension; Indicates spatial dimension; Indicates the channel dimension; Represents matrix multiplication; Three DyHead dynamic detection heads were used to process feature maps with resolutions of 20×20, 40×40, and 80×80 in the Neck network to perform target labeling and category prediction for underground anomalies.