Shield tunnel wall back grouting defect radar image intelligent identification method and system thereof

Abstract

The application discloses a shield tunnel wall behind grouting defect radar image intelligent recognition method and system, belongs to the technical field of shield tunnel nondestructive testing and intelligent image analysis, and comprises the following steps: a radar data acquisition and pretreatment step, background removal and depth attenuation compensation are carried out on a geological radar B-scan image; a multi-scale feature extraction step, encoder multi-scale features, reflection wave amplitude, two-way travel time and waveform envelope physical features are extracted and fused; a wall behind defect semantic segmentation step, four kinds of pixel-level segmentation probability graphs of cavities, non-dense, waterlogging and normal grouting are output; a defect classification and three-dimensional positioning step, image coordinates are converted into tunnel three-dimensional space coordinates; a grouting quality evaluation and report generation step, development maps and positioning reports are generated, and optimization segmentation parameters are fed back, the application adopts semantic segmentation to output pixel-level defect boundaries, multi-source data fusion improves detection accuracy, and a closed-loop feedback mechanism improves detection reliability.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing and intelligent image analysis technology for shield tunnels, specifically relating to a method and system for intelligent identification of grouting defects behind shield tunnel walls using radar images. Background Technology

[0002] Grouting behind the tunnel walls is a crucial process for controlling ground settlement and ensuring the safety of the tunnel structure. During tunnel boring machine (TBM) advancement, the annular gap formed between the outer wall of the tunnel segments and the excavated soil needs to be filled through synchronous grouting or secondary grouting. However, due to factors such as fluctuations in grouting pressure, grout flow properties, and groundwater seepage, defects such as voids, loose areas, and water accumulation areas often appear in the grouting layer. These defects can lead to uneven stress on the tunnel segments and increased ground deformation, seriously threatening the safe operation of the tunnel.

[0003] Ground-penetrating radar (GPR), as a highly efficient non-destructive testing method, transmits high-frequency electromagnetic waves behind the tunnel wall and receives the reflected signals to obtain information about the medium distribution inside the grouting layer. The B-scan image generated by GPR uses the horizontal axis to represent the measurement position along the tunnel direction and the vertical axis to represent the two-way travel time of the electromagnetic waves. The reflected signals from different medium interfaces form waveform responses with distinct characteristics in the image. Cavity areas generate strong reflection signals due to the large difference in dielectric constant between air and grouting materials; loosely compacted areas exhibit chaotic scattering responses due to uneven aggregate distribution and the presence of pores; and waterlogged areas show obvious multiple reflection characteristics due to the high dielectric constant of water.

[0004] In existing technologies, defect detection for ground-penetrating radar (GPR) images mainly employs two types of methods. The first type is image interpretation methods based on human experience, where professional technicians make comprehensive judgments based on features such as reflected wave morphology, amplitude intensity, and two-way travel time. The accuracy of this type of method is highly dependent on the professional level and working conditions of the interpreters, and it is inefficient and highly subjective when dealing with large amounts of data. The second type is target detection methods based on deep learning. For example, Chinese patent CN118628452A discloses a method for detecting internal cavities and non-compact defects in tunnel linings based on GPR data. This method uses an improved DETR deep learning framework, embedding a multi-scale convolutional block attention module into the original architecture, and outputs predicted bounding boxes and class probabilities of defect areas through target detection.

[0005] However, the aforementioned comparative documents have the following technical problems: First, the method targets defects inside the tunnel lining, without considering the special structure and detection requirements of the grouting layer between the shield tunnel segments and the surrounding rock, and the strong reflection interference from the segment reinforcement cannot be effectively eliminated; Second, the target detection method outputs a rectangular prediction box, which cannot accurately describe the actual boundary contour of the defect, and it is difficult to accurately quantify irregularly shaped non-dense areas and water-filled areas; Third, there is a lack of a complete technical solution to convert radar image detection results into the three-dimensional spatial location of defects, and it is impossible to generate an intuitive grouting quality distribution unfolding diagram for engineers to refer to; Fourth, no feedback optimization mechanism for the detection results has been established, and when there are obvious anomalies in the detection results, the segmentation parameters cannot be automatically adjusted to improve the reliability of the detection.

[0006] Therefore, there is an urgent need for an intelligent identification method specifically designed for detecting defects in grouting behind shield tunnel walls, capable of outputting pixel-level segmentation results, providing three-dimensional defect localization and quality assessment reports, and possessing adaptive feedback optimization capabilities. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent identification of grouting defects behind shield tunnel walls using radar images.

[0008] The present invention provides an intelligent radar image recognition method for grouting defects behind shield tunnel walls, comprising the following steps: a radar data acquisition and preprocessing step, acquiring a ground-penetrating radar (GPR) B-scan image of the shield tunnel, associating and matching the GPR B-scan image with segment thickness design data and grouting parameter recording data, and performing background removal and depth attenuation compensation processing on the GPR B-scan image to generate a preprocessed radar image; a multi-scale feature extraction step, inputting the preprocessed radar image into an encoder network for multi-scale downsampling, extracting feature maps at different resolution levels, and simultaneously extracting reflected wave amplitude feature maps, two-way travel time feature maps, and waveform envelope feature maps from the preprocessed radar image, respectively, and concatenating the feature maps at each level with the reflected wave amplitude feature map, two-way travel time feature map, and waveform envelope feature map along the channel dimension to generate a multi-scale fused feature map set; and a wall defect semantic segmentation step, inputting the multi-scale fused feature map set into a decoder network for layer-by-layer upsampling, and fusing through skip connections. The encoder features at the corresponding level output a pixel-level segmentation probability map with the same size as the preprocessed radar image. The pixel-level segmentation probability map contains the probability distribution of four categories: void areas, non-dense areas, water-filled areas, and normal grouting areas. In the defect classification and 3D localization step, the pixel-level segmentation probability map is thresholded to generate masks for various defect areas. The contours of connected defect areas are extracted and centroids are calculated. Based on the mapping relationship between two-way travel time and depth, the image coordinates of the centroids are converted into the 3D spatial coordinates of the defects in the tunnel, generating a set of 3D coordinates of the defects. In the grouting quality assessment and report generation step, the area proportion and distribution density of various defects are statistically analyzed based on the defect area masks. Based on the set of 3D coordinates of the defects, a grouting quality distribution unfolded along the longitudinal and circumferential directions of the tunnel is generated. At the same time, a 3D localization report of the defect location, including the defect type, location, size, and severity, is generated. Based on the consistency of the segmentation results of adjacent detection sections, a segmentation parameter optimization feedback signal is generated and transmitted to the semantic segmentation step of the defect behind the wall.

[0009] This invention also provides an intelligent radar image recognition system for grouting defects behind shield tunnel walls, comprising: a radar data acquisition and preprocessing module for acquiring ground-penetrating radar (GPR) B-scan images of shield tunnels, associating and matching the GPR B-scan images with segment thickness design data and grouting parameter records, performing background removal and depth attenuation compensation processing on the GPR B-scan images to generate preprocessed radar images; a multi-scale feature extraction module for inputting the preprocessed radar images into an encoder network for multi-scale downsampling, extracting feature maps at different resolution levels, and simultaneously extracting reflected wave amplitude feature maps, two-way travel time feature maps, and waveform envelope feature maps to generate a multi-scale fused feature map set; and a semantic segmentation module for behind-wall defects. The module is used to input a multi-scale fused feature map set into the decoder network for layer-by-layer upsampling, outputting a pixel-level segmentation probability map. The pixel-level segmentation probability map contains the probability distribution of four categories: void regions, loose regions, water-filled regions, and normal grouting regions. The defect classification and 3D localization module is used to threshold the pixel-level segmentation probability map to generate a defect region mask. Based on the mapping relationship between two-way travel time and depth, the image coordinates of the defect region are converted into 3D spatial coordinates to generate a set of defect 3D coordinates. The grouting quality assessment and report generation module is used to generate a grouting quality distribution unfolded map and a 3D location report of the defect based on the defect region mask and the defect 3D coordinate set. It also generates a segmentation parameter optimization feedback signal and transmits it to the wall-mounted defect semantic segmentation module.

[0010] The beneficial effects of this invention are as follows: First, by employing semantic segmentation rather than target detection, it can output pixel-level defect boundary contours, accurately quantifying the actual area and shape of cavities, loose areas, and water-filled areas. Second, by fusing radar image features with segment thickness design data and grouting parameter recording data through multi-source data fusion, the targeting and accuracy of detection are improved. Third, by synergistically fusing three types of features—reflected wave amplitude, two-way travel time, and waveform envelope—the ability to distinguish different types of defects is enhanced. Fourth, a mapping mechanism from radar image coordinates to tunnel three-dimensional spatial coordinates is established, automatically generating a three-dimensional location report of defect locations and a grouting quality distribution unfolding diagram. Fifth, a feedback optimization mechanism based on the consistency of adjacent sections is constructed, improving the reliability and stability of detection results by dynamically adjusting segmentation parameters. Actual engineering verification shows that the accuracy rate of cavity detection in this invention reaches over 88%, the accuracy rate of loose area detection reaches over 82%, and the accuracy rate of water-filled area detection reaches over 85%, with overall performance superior to existing technical solutions. Attached Figure Description

[0011] Figure 1 This is a flowchart of the intelligent identification method for radar images of grouting defects behind shield tunnel walls according to the present invention.

[0012] Figure 2This is an architecture diagram of the intelligent radar image recognition system for grouting defects behind shield tunnel walls according to the present invention. Detailed Implementation

[0013] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0014] like Figure 1 As shown, the intelligent identification method for grouting defects behind shield tunnel walls provided by the present invention includes radar data acquisition and preprocessing steps, multi-scale feature extraction steps, semantic segmentation of defects behind the walls steps, defect classification and three-dimensional positioning steps, and grouting quality assessment and report generation steps.

[0015] Step S1: Radar data acquisition and preprocessing.

[0016] The radar data acquisition and preprocessing step is used to acquire and preprocess the ground-penetrating radar detection data of the shield tunnel, providing high-quality input data for subsequent feature extraction and defect segmentation.

[0017] In this embodiment, the ground-penetrating radar (GPR) B-scan images are acquired by moving the GPR antenna longitudinally or circumferentially along the inner wall surface of the shield tunnel segments. The radar equipment uses a shielded antenna with a center frequency of 400MHz to 900MHz, a sampling frequency of 4GHz to 8GHz, and a time window of 50ns to 100ns to cover the detection depth of the segment thickness and the grouting layer behind the wall. The number of scan channels per survey line is determined based on the survey line length and channel spacing, with the channel spacing set to 1cm to 5cm to ensure sufficient lateral resolution.

[0018] The acquired raw radar echo data is stored in the form of a two-dimensional matrix, where the number of rows corresponds to the number of sampling points and the number of columns corresponds to the number of scanning channels. Let the raw radar data matrix be... ,in The number of sampling points for each channel. This represents the total number of scan channels. The raw data contains various signal components, including direct waves, internal reflections of the tunnel segments, reflections from the bottom of the tunnel segments, reflections from the grouting layer, and reflections from the surrounding rock, as well as superimposed system noise and environmental interference.

[0019] Data preprocessing begins with background removal. Background signals primarily originate from direct waves and system coupling signals, which exhibit high correlation across channels. The average channel subtraction method is used to remove background signals: the average waveform of all channels is calculated. Then, the average waveform is subtracted from each signal to obtain the signal after background removal. .

[0020] Depth attenuation compensation is a crucial step in preprocessing. Radar waves propagating in a medium are attenuated by both geometric diffusion and medium absorption, resulting in a much smaller amplitude of the reflected signal from deeper layers compared to shallower layers. This invention employs an adaptive gain compensation algorithm to correct for attenuation. Let the... Channel signal at sampling time The amplitude is The compensated amplitude is:

[0021] ,

[0022] in, The gain function is expressed in exponential form:

[0023] ,

[0024] in, This is the attenuation compensation coefficient, in units of... ; The sampling time is in nanoseconds (ns). The attenuation compensation coefficient is estimated based on the electromagnetic parameters of the grouting material.

[0025] ,

[0026] in, The electrical conductivity of the grouting material is expressed in S / m. A preset value is set according to the grouting mix ratio. The typical value for cement grout is 0.01 S / m to 0.1 S / m. The relative permittivity of the grouting material is typically 6 to 12 for hardened cement grout. Let be the free permeability, and take a value of . H / m; Let be the vacuum permittivity, with a value of . F / m.

[0027] During preprocessing, it is necessary to correlate and match the ground-penetrating radar B-scan images with the segment thickness design data and grouting parameter records. The segment thickness design data contains the design thickness values ​​for each ring of segments, used to determine the time window range within which the reflected signal from the bottom surface of the segment should appear. Let the segment design thickness be... The dielectric constant of the tunnel segment concrete is The theoretical two-way travel time of the signal reflected from the bottom surface of the tube segment is:

[0028] ,

[0029] in, For the speed of light, take a value of m / s; dielectric constant of the tunnel segment concrete Typical values ​​are 6 to 9. The time window is set based on this theoretical value. Used for positioning the bottom surface reflection of the tube segment, among which The tolerance time is set according to the construction deviation of the segment thickness, and the value ranges from 5ns to 10ns.

[0030] The grouting parameter records include construction information such as grouting time, grouting pressure, and grouting volume for each ring segment. The ground-penetrating radar B-scan images are segmented according to the ring number, with each segment corresponding to the detection data of one ring segment. When the grouting parameter records show an abnormal drop in grouting pressure or insufficient grouting volume in a certain ring, a preliminary anomaly marker is added to the radar data segment corresponding to that ring, prompting subsequent segmentation steps to focus on this area.

[0031] After background removal and depth attenuation compensation, a preprocessed radar image is generated. To facilitate subsequent deep learning network processing, the preprocessed two-dimensional data matrix is ​​converted into a three-channel image format, with each channel storing the original amplitude, absolute amplitude value, and amplitude polarity information, respectively. The image size is normalized to... ,in The corresponding number of sampling points on the time axis, with values ​​of 256 or 512; The number of scan channels corresponding to the spatial axis is determined based on the actual measurement line length.

[0032] Reinforcing bar signal suppression is also required during preprocessing. Shield tunnel segments typically contain circumferential main reinforcement and longitudinal distribution reinforcement. As strong reflectors, these reinforcements generate noticeable hyperbolic clutter in radar images, interfering with the identification of grouting defects behind the tunnel wall. This invention employs a morphological filtering-based reinforcing bar signal suppression method: First, the preprocessed radar image undergoes median filtering in the column direction. The filtering window size is set according to the reinforcement spacing; when the reinforcement spacing is 15cm to 20cm and the channel spacing is 2cm, the window size is set to 9 to 11 channels. Then, the median filtering result is subtracted from the original image to obtain an enhanced reinforcing bar signal image. Finally, the enhanced reinforcing bar signal image is used to adaptively suppress the original image, with the suppression intensity automatically adjusted according to the signal energy.

[0033] In addition, the preprocessing steps include time zero-point correction and bandpass filtering. Time zero-point correction is used to eliminate time offsets caused by direct waves and system delays, achieved by detecting the first arrival position of each channel signal and aligning it to a unified reference time. Bandpass filtering is used to remove high-frequency random noise and low-frequency baseline drift, employing a Butterworth bandpass filter with the lower cutoff frequency set to 0.3 times the antenna center frequency and the upper cutoff frequency set to 1.5 times the antenna center frequency.

[0034] Step S2: Multi-scale feature extraction step.

[0035] The multi-scale feature extraction step is used to extract multi-scale semantic features and multi-type physical features from the preprocessed radar image, providing rich feature representations for subsequent semantic segmentation.

[0036] The encoder network in this step adopts the ResNet series residual network structure, and ResNet-50 is selected as the backbone network in this embodiment. After the preprocessed radar image is input into the encoder network, it passes through the initial convolutional layer and the max pooling layer, and then sequentially passes through four residual block groups for multi-scale downsampling.

[0037] Let the size of the preprocessed radar image be... The initial convolutional layer uses The convolution kernel has a stride of 2, and the output feature map size is [size missing]. The max pooling layer uses Pooling kernel with a stride of 2, output feature map size is .

[0038] The first residual block group contains 3 residual blocks, and the output feature map size is [size missing]. The downsampling factor is 4. The second residual block group contains 4 residual blocks, and the output feature map size is... The downsampling factor is 8. The third residual block group contains 6 residual blocks, and the output feature map size is... The downsampling factor is 16. The fourth residual block group contains 3 residual blocks, and the output feature map size is... The downsampling factor is 32.

[0039] To adapt to the task of detecting defects in grouting behind the wall, channel compression convolution was added after each residual block group, compressing the number of feature map channels to 64, 128, 256, and 512 respectively, in order to reduce computational load and highlight features relevant to defect detection. The compressed feature map is denoted as... This forms the feature pyramid of the multi-scale encoder.

[0040] Simultaneously with encoder feature extraction, this step also extracts three types of physical features from the preprocessed radar image: reflected wave amplitude feature map, two-way travel time feature map, and waveform envelope feature map. These three types of features directly reflect the physical characteristics of the radar signal and have a clear correspondence with the radar response of different types of defects.

[0041] The method for extracting the reflected wave amplitude feature map is as follows: For the first... channel signal Calculate the absolute mean and peak The absolute mean and peak values ​​of each channel are used to create two one-dimensional images, which are then copied and expanded along the time axis to form two-dimensional feature maps of the same size as the preprocessed radar image, thus obtaining the amplitude mean feature map. and amplitude peak characteristic map The two feature maps are stitched together along the channel dimension to form the reflected wave amplitude feature map. .

[0042] The method for extracting the two-way travel time feature map is as follows: For the first... The arrival time of the first wave is determined by the energy arrival detection algorithm. and the moment of the main reflected wave The first arrival time is defined as the moment when the signal energy first exceeds the noise floor. Sampling time times: ,in The standard deviation of the signal at the first few sampling points. The threshold coefficient is set to a value between 3 and 5. The moment of the main reflection wave is defined as the peak moment of the reflected signal from the bottom surface of the segment to the grouting layer interface, which is determined by searching for the position of the maximum amplitude within the time window on the bottom surface of the segment. The arrival time of the first wave and the moment of the main reflection wave of each trace are respectively composed into two one-dimensional images and expanded into two-dimensional feature maps, which are then stitched together to obtain a two-way travel time feature map. .

[0043] The method for extracting the waveform envelope feature map is as follows: For the th channel signal Analytical signals are calculated using Hilbert transform. ,in For Hilbert transform operators, The unit is imaginary. Instantaneous amplitude (envelope) The envelope signals of each channel are combined to form a two-dimensional image to obtain the waveform envelope feature map. .

[0044] After the three types of physical feature maps are compressed and convolved to unify the number of channels, they are then fused with feature maps from each level of the encoder using multi-feature adaptive fusion. The fusion weights are adaptively calculated using an attention mechanism. Let the encoder's... Layer feature map is The three types of physical feature maps are downsampled to the same size and then stitched together. The fused feature map is then calculated as follows:

[0045] ,

[0046] in, for The convolution operation is used to align the number of channels in the physical feature map to the number of channels in the encoder feature map; For spatial attention weights, by... It is obtained by performing global average pooling, fully connected layers, and Sigmoid activation calculations; This indicates element-wise multiplication.

[0047] The formula for calculating the fusion weight is:

[0048] ,

[0049] in, For global average pooling, the feature map is compressed into a vector with channel dimensions; and It consists of two fully connected layers, with the number of intermediate channels compressed to half the original number. ; To modify the activation function of the linear unit; The Sigmoid activation function maps the weight values ​​to... Interval.

[0050] After multi-feature fusion, a multi-scale fused feature map set is generated. This serves as the input for subsequent semantic segmentation steps.

[0051] In the multi-scale feature extraction process, this invention also introduces a spatial pyramid pooling module to enhance the network's ability to perceive defects of different sizes. The spatial pyramid pooling module performs multi-scale pooling on the deepest feature map of the encoder, with the pooling window size set to... , , and The pooling results are then upsampled back to their original size and concatenated with the original feature map. This design enables the network to capture both local detail features and global contextual information, demonstrating good performance in detecting both large-scale non-dense regions and small-sized voids and defects.

[0052] Residual connections and batch normalization strategies were also employed during feature extraction to accelerate network training and prevent gradient vanishing. Each convolutional layer was followed by a batch normalization layer and a ReLU activation function, and the residual blocks used a method with... The convolutional identity mapping matches the number of input and output channels. Network parameters are initialized using the He initialization method, with weights following a mean of 0 and a variance of 0. The normal distribution, where This is the number of input channels.

[0053] Step S3: Semantic segmentation of defects behind the wall.

[0054] The semantic segmentation step behind the wall is used to classify the defect type of radar image at the pixel level, and outputs a pixel-level segmentation probability map containing the probability distribution of four categories: void area, loose area, water accumulation area and normal grouting area.

[0055] The decoder network in this step employs a symmetrical upsampling structure, gradually restoring the spatial resolution of the feature maps through layer-by-layer upsampling and skip connections. The decoder contains four upsampling levels, each of which enlarges the feature map size by a factor of 2 through bilinear interpolation and then concatenates the features with those from the encoder at the corresponding level via skip connections.

[0056] Let the decoder be the first The input feature map of the layer is The corresponding encoder fused feature map is Then the decoder's first The output feature map of the layer is calculated as follows:

[0057] ,

[0058] in, For bilinear interpolation upsampling, the feature map size is increased by a factor of 2; For the attention gating module, the encoder features are weighted and filtered based on the decoder features; This is a channel-level splicing operation; for Convolutional layer.

[0059] The attention gating module is one of the key innovations of this invention, used to adaptively filter encoder features related to the current decoding level. The calculation formula for attention gating is: ,

[0060] Two of them The layers map the encoder features and the upsampled decoder features to the same number of intermediate channels, respectively. To correct the activation function of the linear unit; the last one The layer maps the activated features to a single-channel attention weight map; The sigmoid activation function is used. The encoder features after attention gating are... .

[0061] After four upsampling levels, the decoder output feature map size is restored to the same size as the preprocessed radar image. Connect after the decoder output. Convolutional layers and softmax activation layers convert the feature map into a four-channel pixel-level segmentation probability map. ,in This represents the number of categories.

[0062] Each pixel position in the pixel-level segmentation probability map Corresponding to one 3D probability vector The four components represent the probability that the pixel belongs to a void region, a loose region, a water-filled region, and a normal grouting region, respectively, satisfying the following conditions: The predicted category for each pixel is the category with the highest probability. .

[0063] The segmentation network is trained using a combined loss function, consisting of cross-entropy loss and Dice loss. Let the ground truth labels be... The combined loss function is:

[0064] ,

[0065] in, and As the loss weighting coefficient, this embodiment sets and .

[0066] The cross-entropy loss is calculated as follows:

[0067] ,

[0068] in, For the one-hot encoding of the true label, when the true category is hour ,otherwise ; To prevent overflow in logarithmic operations, a small constant is selected. .

[0069] Because the number of samples in each category of the backfill grouting defect dataset is imbalanced, with far more samples in the normal grouting area than in the defect area, this invention employs class-weighted Dice loss to alleviate the class imbalance problem. The Dice loss is calculated as follows:

[0070] ,

[0071] in, For the first The loss weights for each class are set based on the reciprocal of the proportion of each class in the training set. Let the first class be the loss weight. The number of samples in each class is The total number of samples is ,but .

[0072] During the inference phase, segmentation confidence is calculated for each pixel location in the pixel-level segmentation probability map. Segmentation confidence is defined as the difference between the predicted class probability and the probability of the second-highest class.

[0073] ,

[0074] When the segmentation confidence level is lower than the preset confidence threshold When this occurs, the pixel is marked as an area requiring manual review. The preset confidence threshold ranges from 0.7 to 0.9, and is set according to the actual engineering requirements for detection accuracy and false negative rate. Setting a lower threshold can reduce the area to be reviewed but may increase the risk of false negatives, while setting a higher threshold can improve detection reliability but increase the workload of manual review.

[0075] During the training of the segmentation network, this invention employs data augmentation strategies to expand the diversity of training samples and improve the model's generalization ability. Data augmentation operations include: random horizontal flipping with a probability of 0.5; random vertical flipping with a probability of 0.3; random rotation with an angle range of ±10 degrees; random scaling with a scaling ratio range of 0.8 to 1.2; random brightness adjustment with an adjustment range of ±20%; and adding Gaussian noise with a noise standard deviation range of 0 to 0.02. Data augmentation is performed online during training, with each training sample randomly applying a combination of the above augmentation operations in each training round.

[0076] To address the issue of scarce defect samples in the training dataset, this invention also employs a radar simulation data generation method based on the finite-difference time-domain method. A finite element model of the tunnel cross-section, including segments, grouting layer, and surrounding rock, is established using open-source software such as gprMax. Cavities, loose areas, and water-filled areas of different sizes and locations are defined, and then forward modeling of the radar wavefield is performed to generate synthetic B-scan images. The mixing ratio of simulation data to measured data is set to 1:3, with the simulation data primarily used for model warm-up in the initial training phase and for supplementing samples of rare defect types.

[0077] The network was trained using the Adam optimizer with an initial learning rate of 0.001 and a weight decay factor of 0.0001. The learning rate was dynamically adjusted using a cosine annealing strategy, decaying from its initial value to a minimum of 0.00001 within each training epoch according to a cosine function before resuming. The training batch size was set based on the GPU memory capacity, set to 8 on an NVIDIA RTX3090 GPU and 16 on an NVIDIA A100 GPU. An early stopping strategy was employed during training; training was terminated when the validation set loss did not decrease for 20 consecutive epochs, and the model parameters with the best performance on the validation set were selected as the final model.

[0078] Step S4: Defect classification and 3D localization steps.

[0079] The defect classification and 3D localization steps are used to post-process the semantic segmentation results, extract the shape and location information of various defect regions, and convert the image coordinates into tunnel 3D spatial coordinates.

[0080] First, the pixel-level segmentation probability map is thresholded to generate masks for various defect regions. Let the segmentation threshold be... If the value is 0.5, then the first... The defect-like region mask is calculated as follows:

[0081] ,

[0082] for The corresponding hole area masks Non-dense area mask and water accumulation area mask Post-processing was performed using morphological operations. First, using… The erosion kernel is used to erode the mask to remove isolated noise points, and then... The expanding core undergoes an expansion operation to fill the small holes.

[0083] Next, connected component analysis is performed on the masks of various defect regions to extract connected defect regions and calculate their geometric features. For the first... Defect-like region mask The 8-connected region labeling algorithm is used to identify all connected components, denoted as . ,in For the first The number of connected components in a defective region.

[0084] For each connected component Calculate the following geometric features: area The number of pixels within the connected region; perimeter. The number of pixels at the boundary of the connected domain; centroid The mean of all pixel coordinates within the connected component; the bounding rectangle. The smallest axis-aligned rectangle enclosing the connected region; circularity Used to describe the degree of regularity of a shape.

[0085] Converting image coordinates to tunnel 3D spatial coordinates is the core task of this step. Shield tunnels are represented using a cylindrical coordinate system. These represent the position along the tunnel axis, the circumferential angle, and the radial depth, respectively.

[0086] In the image coordinate system, the horizontal axis Corresponding to the survey line direction, the vertical axis Corresponding to the two-way travel time of radar waves. For survey lines arranged longitudinally along the tunnel, the horizontal axis corresponds to the tunnel axis direction. For survey lines arranged circumferentially along the tunnel, the horizontal axis corresponds to the circumferential angle. Let the starting point of the survey line be located in the tunnel coordinate system as follows: Lane spacing is Then the horizontal axis coordinate of the image The corresponding tunnel location is:

[0087] Longitudinal survey line: , ;

[0088] Circumferential survey line: , ;

[0089] in, This refers to the inner diameter of the tunnel.

[0090] Image vertical axis coordinates Corresponding radar wave two-way travel time ,in The sampling time interval is given. The mapping relationship between two-way travel time and depth needs to be calculated in segments based on the dielectric constants of the segments and grouting materials.

[0091] Let the thickness of the tunnel segment be The dielectric constant of the tube segment is The two-way travel time of the bottom surface reflection of the tube segment is For two-way travel time The area corresponding to the interior of the tunnel segment has its depth calculated as follows:

[0092] ,

[0093] For two-way travel time The area corresponding to the backfill grouting layer and deeper locations is calculated as follows:

[0094] ,

[0095] in, This represents the dielectric constant of the grouting material. In practical applications, the dielectric constant can be obtained through reflection calibration on the bottom surface of the lining segment or through on-site core drilling tests.

[0096] For the center of mass located at The The first The three-dimensional spatial coordinates of the defect-like region are calculated as follows: The three-dimensional coordinates of all defect areas constitute the set of three-dimensional coordinates of the defect:

[0097] ,

[0098] Each element contains the defect's location coordinates, type code, and area information.

[0099] For estimating the size of the defect area, in addition to calculating the pixel area, it is also necessary to convert the pixel area into the actual physical area. Let the physical resolution in the horizontal axis be... (Unit: m / pixel), calculated based on channel spacing and image size; physical resolution along the vertical axis. Calculated based on the sampling time interval and radar wave propagation speed. The physical area of ​​the defect region. The calculation is as follows: ,in The pixel area.

[0100] For irregularly shaped defect areas, this step also calculates other geometric features to support subsequent defect severity assessment. Major axis length. and minor axis length The extension direction is obtained by fitting the contour of the defect area with a minimum bounding ellipse. The angle between the major axis of the smallest circumscribed ellipse and the horizontal direction; eccentricity. Used to describe the flatness of the defect shape. Void defects are usually elliptical or circular with low eccentricity; loose areas are often distributed in irregular strips or sheets with high eccentricity.

[0101] To address potential inaccuracies in segmentation boundaries, this step employs a Conditional Random Field (CRF) model for post-processing optimization of the segmentation results. The CRF model considers the spatial relationships between adjacent pixels and class consistency constraints, performing global optimization of the segmentation probability map. The energy function includes unary and binary potential terms: the unary term is defined based on the class probabilities output by the segmentation network; the binary term uses a Gaussian kernel function, encouraging adjacent pixels to be segmented into the same class when they are close in image space and have similar grayscale values. The CRF inference is solved iteratively using a mean-field approximation algorithm, with 10 iterations.

[0102] Step S5: Grouting quality assessment and report generation.

[0103] The grouting quality assessment and report generation step is used to generate a visualized grouting quality distribution map and a structured 3D location report of defects based on the defect detection results, and to optimize the segmentation parameters through a feedback mechanism.

[0104] The grouting quality distribution diagram unfolds the surface of the cylindrical tunnel into a two-dimensional planar view, with the horizontal axis representing the tunnel axis direction. The vertical axis represents the circumferential angle. Or unfold the arc length The resolution of the unfolded image is set according to the detection accuracy requirements. In this embodiment, the horizontal axis resolution is 10cm and the vertical axis resolution is 5°.

[0105] The unfolded diagram uses color coding to represent the grouting quality status at each location. Let the unfolded diagram location... The quality rating at the location is The comprehensive scoring algorithm for grouting quality is calculated based on the presence and type of defects at that location.

[0106] ,

[0107] in, For the first The projection area of ​​the defect-like region onto the unfolded diagram; This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. For the first Severity weights of class defects, voids Not dense Floodwater .

[0108] Grouting quality score The range of values ​​is A value of 1 indicates that the grouting quality at that location is normal, while a lower value indicates a more severe defect. The unfolded diagram is shown in green. Excellent areas, indicated by yellow The general area, indicated by red. The defective area.

[0109] The 3D defect location report is output in a structured format and includes the following: basic inspection information (inspection date, inspection mileage range, inspection equipment parameters); defect statistics (number of various defects, total area, area percentage, distribution density); detailed list of defects (type, location coordinates, size, severity level of each defect area); and quality assessment conclusion (overall quality level, list of defects requiring treatment, and treatment recommendations).

[0110] The severity of a defect is determined based on its type and size. For void defects, areas less than 0.1 square meters are considered minor, 0.1 to 0.5 square meters are moderate, and larger than 0.5 square meters are severe. For loose defects, areas less than 0.2 square meters are considered minor, 0.2 to 1.0 square meters are moderate, and larger than 1.0 square meters are severe. For water accumulation defects, areas less than 0.05 square meters are considered minor, 0.05 to 0.2 square meters are moderate, and larger than 0.2 square meters are severe.

[0111] In addition to the severity of individual defects, this step also calculates grouting quality indicators at the section level. For a tunnel section of a specified length, the following indicators are calculated: defect coverage rate is the percentage of the total area of ​​all defective areas within the section to the total area of ​​the inspected area; defect density is the number of defects per unit length of tunnel; the defect distribution uniformity index is obtained by calculating the coefficient of variation of the defect coverage rate of each ring segment. A larger coefficient of variation indicates a more uneven defect distribution, which may indicate severe local deterioration in grouting quality.

[0112] Based on the above indicators, this step establishes a grouting quality grading standard. Grade 1 (Excellent): Defect coverage less than 5% and no serious defects; Grade 2 (Good): Defect coverage 5% to 15% or moderate defects but no serious defects; Grade 3 (Medium): Defect coverage 15% to 30% or serious defects but no more than 2; Grade 4 (Poor): Defect coverage exceeding 30% or more than 2 serious defects. Different treatment recommendations correspond to different grades: Grade 1 requires no treatment; Grade 2 recommends enhanced monitoring; Grade 3 recommends supplementary grouting; Grade 4 recommends immediate treatment and investigation of the cause.

[0113] The segmentation parameter optimization feedback mechanism is a key innovation of this invention, which automatically adjusts the segmentation parameters by analyzing the spatiotemporal consistency of the detection results. For multiple detection sections continuously collected along the tunnel longitudinal direction, the defect type and size of adjacent sections at the same circumferential position should have high consistency, because actual defects usually extend for a certain length along the longitudinal direction.

[0114] Assume adjacent detection sections and In the circumferential position The defect types at each location are respectively and Define consistency metrics:

[0115] ,

[0116] When the consistency index is 0, it indicates a sudden change in defect identification results between adjacent sections, possibly due to insufficient sensitivity of the segmentation parameters leading to missed detections or excessive sensitivity leading to false detections. At this point, a segmentation parameter optimization feedback signal is generated: if the section... The cross section was identified as a defect. If identified as normal, the cross-section is reduced. The segmentation threshold at this location The reduction ranges from 0.05 to 0.1; if the cross-section Identified as normal cross-section If a defect is identified, the two cross-sections at that location are re-segmented, and the classification result with the higher probability value is selected.

[0117] Furthermore, when the defect area ratio of multiple consecutive detection sections (the number exceeding a preset threshold for the number of consecutive sections, ranging from 3 to 5) exceeds a preset area ratio threshold (ranging from 15% to 25%), it indicates that there is a large-area grouting defect in that section, requiring an increase in the feature fusion weights of the segmentation network to enhance the defect boundary recognition capability. Specifically, this involves adjusting the attention weights in the multi-feature fusion weighting formula. Multiply by the enhancement factor , The value ranges from 1.2 to 1.5.

[0118] like Figure 2 As shown, the intelligent radar image recognition system for grouting defects behind the shield tunnel wall provided by the present invention includes a radar data acquisition and preprocessing module 1, a multi-scale feature extraction module 2, a semantic segmentation module for defects behind the wall 3, a defect classification and three-dimensional positioning module 4, and a grouting quality assessment and report generation module 5.

[0119] The radar data acquisition and preprocessing module 1 is used to acquire ground-penetrating radar (GPR) B-scan images of the shield tunnel. It correlates and matches these GPR B-scan images with segment thickness design data and grouting parameter records, performing background removal and depth attenuation compensation on the GPR B-scan images to generate preprocessed radar images. This module receives raw echo data from the GPR equipment, segment thickness design data from the design drawing database, and grouting parameter records from the construction record system. The preprocessed radar images output by this module are then passed to the multi-scale feature extraction module. The adaptive gain compensation algorithm, background removal method, and data correlation matching mechanism used in the preprocessing process are the same as those described in the method embodiment.

[0120] The multi-scale feature extraction module 2 is used to input the preprocessed radar image into the encoder network for multi-scale downsampling, extracting feature maps at different resolution levels, and simultaneously extracting reflected wave amplitude feature maps, two-way travel time feature maps, and waveform envelope feature maps to generate a multi-scale fused feature map set. This module receives the preprocessed radar image from the radar data acquisition and preprocessing module 1 and outputs the multi-scale fused feature map set to be passed to the wall-back defect semantic segmentation module 3. The encoder network structure, physical feature extraction method, and multi-feature fusion mechanism are the same as those described in the method embodiment.

[0121] The semantic segmentation module 3 for wall defects is used to input the multi-scale fused feature map set into the decoder network for layer-by-layer upsampling, outputting a pixel-level segmentation probability map. This pixel-level segmentation probability map contains the probability distributions for four categories: void regions, loose regions, water-filled regions, and normal grouting regions. This module receives the multi-scale fused feature map set from the multi-scale feature extraction module 2 and the segmentation parameter optimization feedback signal from the grouting quality assessment and report generation module 5, and outputs the pixel-level segmentation probability map to the defect classification and 3D localization module. The decoder network structure, attention gating mechanism, and loss function design and method are the same as described in the previous embodiment. The segmentation parameter optimization feedback signal is used to dynamically adjust the segmentation threshold and feature fusion weights.

[0122] The defect classification and 3D localization module 4 is used to threshold the pixel-level segmentation probability map to generate a defect region mask. Based on the mapping relationship between two-way travel time and depth, it converts the image coordinates of the defect region into 3D spatial coordinates, generating a set of 3D defect coordinates. This module receives the pixel-level segmentation probability map from the wall-mounted defect semantic segmentation module 3 and outputs the defect region mask and the set of 3D defect coordinates, which are then passed to the grouting quality assessment and report generation module 5. The thresholding, morphological post-processing, connected component analysis, and coordinate transformation methods are the same as those described in the method embodiments.

[0123] The grouting quality assessment and report generation module 5 generates a grouting quality distribution unfolded map and a 3D location report of defects based on the defect area mask and the set of 3D coordinates of the defects. It also generates a segmentation parameter optimization feedback signal, which is transmitted to the back-wall defect semantic segmentation module 3. This module receives the defect area mask and the set of 3D coordinates of defects from the defect classification and 3D location module 4, outputs the grouting quality distribution unfolded map and the 3D location report of defects as the final output of the system, and simultaneously generates a segmentation parameter optimization feedback signal, which is fed back to the back-wall defect semantic segmentation module 3 to form a closed loop. The quality scoring algorithm, unfolded map generation method, report format, and feedback optimization mechanism are the same as those described in the method embodiment.

[0124] The five modules of this invention form a deeply coupled closed-loop collaborative architecture. The output of the radar data acquisition and preprocessing module 1 is the sole input source for the multi-scale feature extraction module, while the output of the multi-scale feature extraction module 2 is the primary input source for the back-wall defect semantic segmentation module 3. The output of the back-wall defect semantic segmentation module 3 is processed by the defect classification and 3D localization module 4 and then transmitted to the grouting quality assessment and report generation module 5. The segmentation parameter optimization feedback signal generated by the latter is then transmitted back to the back-wall defect semantic segmentation module 3, enabling dynamic adjustment of the segmentation parameters based on the detection results. This combined feedforward and feedback architecture allows the system to adaptively optimize the segmentation effect according to the actual detection situation, improving the reliability and stability of the detection results.

[0125] From a system deployment perspective, this invention supports two deployment modes: edge computing and cloud computing. In edge computing mode, the system is deployed on an industrial computer at the tunnel construction site, configured with a GPU accelerator card for real-time detection. Detection latency is controlled to within 2 seconds for a single B-scan image, suitable for detection scenarios requiring immediate on-site feedback. In cloud computing mode, the system is deployed on a remote server cluster, receiving radar data collected on-site via a network and returning detection results. This mode utilizes more powerful computing resources to process large volumes of data, suitable for detection scenarios requiring post-event batch analysis.

[0126] The system's user interface is designed with simplicity and intuitiveness in mind. The main interface comprises four functional areas: a data import area, a detection parameter setting area, a real-time preview area, and a result output area. The data import area supports both single-file and batch import modes, and supports common radar data formats such as SEG-Y and DZT. The detection parameter setting area provides configuration options for key parameters such as segment parameters, grouting parameters, and detection sensitivity, and includes preset parameter templates for various typical working conditions. The real-time preview area displays the segmentation probability map and defect detection results in heatmap format. The result output area provides functions such as report generation, unfolded diagram export, and data archiving.

[0127] To verify the effectiveness of the method of this invention, a field test was conducted in a subway shield tunnel project in a certain city. The test tunnel had an outer diameter of 6.2m, an inner diameter of 5.5m, and a segment thickness of 0.35m. The grouting material was cement-fly ash slurry. A ground-penetrating radar with a shielded antenna and a center frequency of 800MHz was used, and survey lines were arranged along the longitudinal and circumferential directions of the tunnel, collecting a total of 1200 B-scan images.

[0128] The training dataset contains 800 labeled B-scan images, with 12,350 pixels labeled for hollow regions, 28,760 pixels for loose regions, 8,920 pixels labeled for waterlogged regions, and 4,125,970 pixels labeled for normal grouting regions. The validation dataset contains 200 images, and the test dataset contains 200 images. The segmentation network is implemented using the PyTorch framework and trained on an NVIDIA RTX 3090 GPU. The batch size is set to 8, the initial learning rate is set to 0.001 with cosine annealing at decay, and the number of training epochs is set to 200.

[0129] Test results show that the method of this invention achieves an accuracy of 88.3% for hole detection, a recall of 85.7%, and an F1 score of 86.9%; an accuracy of 82.5% for non-dense area detection, a recall of 79.8%, and an F1 score of 81.1%; an accuracy of 85.2% for water accumulation area detection, a recall of 82.4%, and an F1 score of 83.8%; and an overall pixel-level segmentation accuracy (mIoU) of 76.8%. Compared with the target detection method used in CN118628452A, the method of this invention significantly improves both the accuracy of defect boundary recognition and the ability to detect small targets.

[0130] To further evaluate the practical performance of the method of this invention, the detection efficiency and accuracy of the method of this invention were compared and analyzed with those of manual interpretation methods and traditional target detection methods. Manual interpretation was performed by professional technicians with more than 5 years of experience in tunnel radar detection, with an average processing time of approximately 8 minutes per B-scan image. Traditional target detection methods used the YOLOv5 model, trained on the same dataset, and then tested. The results show that the processing time of the method of this invention per image is approximately 1.2 seconds, far superior to manual interpretation. Compared with the YOLOv5 target detection method, the method of this invention improves the accuracy of defect boundary contours by approximately 15% and the recall rate for small-sized defects (less than 0.02 square meters) by approximately 22%.

[0131] Furthermore, the generalization ability of this invention was also tested on another urban subway tunnel and a hydraulic tunnel. The urban subway tunnel used different segment thicknesses and grouting materials, while the hydraulic tunnel's detection environment had high electromagnetic interference. The test results showed that, without model retraining, the method of this invention achieved an mIoU of 71.2% on the urban subway tunnel and 68.5% on the hydraulic tunnel; after fine-tuning with a small number of samples, the performance improved to 74.8% and 72.1% respectively, demonstrating that the method of this invention has good cross-scenario generalization ability.

[0132] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for intelligent identification of grouting defects behind shield tunnel walls using radar images, characterized in that, Includes the following steps: The radar data acquisition and preprocessing steps involve acquiring the B-scan image of the shield tunnel using ground radar, matching the B-scan image with the segment thickness design data and grouting parameter recording data, removing background and compensating for depth attenuation in the B-scan image, and generating a preprocessed radar image. The multi-scale feature extraction step involves inputting the preprocessed radar image into an encoder network for multi-scale downsampling to extract feature maps at different resolution levels. Simultaneously, the reflected wave amplitude feature map, two-way travel time feature map, and waveform envelope feature map are extracted from the preprocessed radar image. The feature maps at each level are then concatenated with the reflected wave amplitude feature map, two-way travel time feature map, and waveform envelope feature map along the channel dimension to generate a multi-scale fused feature map set. The semantic segmentation step for defects behind the wall involves inputting the multi-scale fused feature map set into the decoder network for layer-by-layer upsampling, fusing the encoder features of the corresponding layers through skip connections, and outputting a pixel-level segmentation probability map with the same size as the preprocessed radar image. The pixel-level segmentation probability map contains the probability distribution of four categories: void area, loose area, water accumulation area, and normal grouting area. The defect classification and 3D localization steps involve thresholding the pixel-level segmentation probability map to generate various defect region masks, extracting the contours and calculating the centroids of connected defect regions, and converting the image coordinates of the centroids into the 3D spatial coordinates of the defects in the tunnel based on the mapping relationship between two-way travel time and depth, thereby generating a set of 3D coordinates of the defects.

2. The intelligent identification method for radar images of grouting defects behind shield tunnel walls according to claim 1, characterized in that, In the radar data acquisition and preprocessing steps, the depth attenuation compensation process includes: performing exponential gain compensation on the amplitude values ​​of the ground-penetrating radar B-scan image at each sampling time based on the attenuation coefficient of the radar wave in the grouting material. The attenuation coefficient is calculated based on the estimated dielectric constant of the grouting material and the preset conductivity parameter.

3. The intelligent identification method for radar images of grouting defects behind shield tunnel walls according to claim 1, characterized in that, In the radar data acquisition and preprocessing steps, the segment thickness design data is used to determine the time window range of the bottom surface reflection signal of the segment in the ground-penetrating radar B-scan image, and the grouting parameter recording data is used to perform data segmentation by ring number and preliminary marking of grouting abnormal areas in the ground-penetrating radar B-scan image.

4. The intelligent identification method for radar images of grouting defects behind shield tunnel walls according to claim 1, characterized in that, In the multi-scale feature extraction step, the encoder network contains four downsampling levels, with downsampling ratios of 4, 8, 16 and 32 for each level, corresponding to 64, 128, 256 and 512 channels of the extracted feature maps, respectively.

5. The intelligent identification method for radar images of grouting defects behind shield tunnel walls according to claim 1, characterized in that, In the semantic segmentation step of the wall-back defect, the segmentation confidence score is calculated for each pixel position in the pixel-level segmentation probability map. When the segmentation confidence score is lower than a preset confidence threshold, the pixel is marked as a region to be manually reviewed. The preset confidence threshold ranges from 0.7 to 0.

9.

6. The intelligent identification method for radar images of grouting defects behind shield tunnel walls according to claim 1, characterized in that, In the multi-scale feature extraction step, the reflected wave amplitude feature map is obtained by calculating the absolute mean and peak value of each channel signal of the preprocessed radar image; the two-way travel time feature map is obtained by detecting the first arrival time and the main reflection time of each channel signal; and the waveform envelope feature map is obtained by performing Hilbert transform on each channel signal and calculating the instantaneous amplitude.

7. The intelligent identification method for radar images of grouting defects behind shield tunnel walls according to claim 1, characterized in that, In the semantic segmentation step of the behind-the-wall defect, the decoder network uses an attention gating mechanism to weight and filter the encoder features passed by the skip connections. The attention gating mechanism calculates the spatial attention weights based on the correlation between the current layer features of the decoder and the corresponding encoder layer features.

8. The intelligent identification method for radar images of grouting defects behind shield tunnel walls according to claim 1, characterized in that, In the defect classification and three-dimensional positioning steps, the mapping relationship between the two-way travel time and the depth is calculated based on the position of the reflected signal on the bottom surface of the segment calibrated by the segment thickness design data and the preset dielectric constant value of the grouting material to obtain the propagation speed of the radar wave in the grouting layer, and then the two-way travel time is converted into a depth value.

9. The intelligent identification method for radar images of grouting defects behind shield tunnel walls according to claim 1, characterized in that, It also includes a grouting quality assessment and report generation step, which calculates the area ratio and distribution density of various defects based on the defect area mask, generates a grouting quality distribution unfolding map along the longitudinal and circumferential directions of the tunnel based on the defect three-dimensional coordinate set, generates a three-dimensional location report of defects including defect type, location, size and severity, and generates a segmentation parameter optimization feedback signal based on the consistency of the segmentation results of adjacent detection sections, which is then transmitted to the wall-back defect semantic segmentation step. The segmentation parameter optimization feedback signal includes: when the defect type identification results at the same location of adjacent detection sections are inconsistent, reducing the segmentation threshold at that location to improve detection sensitivity; When the proportion of defect area in multiple consecutive detection sections exceeds the preset area proportion threshold, the feature fusion weight of the segmentation network is increased to enhance the defect boundary recognition capability.

10. A radar image intelligent recognition system for grouting defects behind shield tunnel walls, used to implement the radar image intelligent recognition method for grouting defects behind shield tunnel walls as described in claim 9, characterized in that, include: The radar data acquisition and preprocessing module is used to acquire the B-scan image of the shield tunnel, associate and match the B-scan image with the segment thickness design data and grouting parameter record data, and perform background removal and depth attenuation compensation processing on the B-scan image to generate a preprocessed radar image. The multi-scale feature extraction module is used to input the preprocessed radar image into the encoder network for multi-scale downsampling, extract feature maps at different resolution levels, and simultaneously extract the reflected wave amplitude feature map, two-way travel time feature map and waveform envelope feature map to generate a multi-scale fused feature map set. The semantic segmentation module for defects behind the wall is used to input the multi-scale fused feature map set into the decoder network for layer-by-layer upsampling and output a pixel-level segmentation probability map. The pixel-level segmentation probability map contains the probability distribution of four categories: void area, loose area, water accumulation area and normal grouting area. The defect classification and 3D localization module is used to threshold the pixel-level segmentation probability map to generate a defect region mask, and convert the image coordinates of the defect region into 3D spatial coordinates according to the mapping relationship between two-way travel time and depth to generate a set of 3D coordinates of the defect. The grouting quality assessment and report generation module is used to generate a grouting quality distribution unfolding map and a three-dimensional positioning report of the defect location based on the defect area mask and the defect three-dimensional coordinate set, and to generate a segmentation parameter optimization feedback signal to be transmitted to the wall-back defect semantic segmentation module.

Images