Ground penetrating radar detection system and method for slurry balance shield

Abstract

The application discloses a ground penetrating radar detection system and method for a slurry balance shield, and relates to the technical field of intelligent monitoring of tunnel engineering. It comprises: injecting Gaussian white noise and power frequency interference into a geological numerical model to obtain a training data set and train a multi-branch decoding network; using cutter position data to perform spatial correction and normalization on real-time echoes, inputting the processed images into a model inversion to obtain a stratum structure probability graph; and searching for a maximum probability index, identifying connectivity, and counting pixels to obtain a slurry film thickness. The application realizes in-situ, real-time and non-destructive monitoring of the quality of the slurry film during construction, solves the problems of traditional methods relying on indirect inference or shutdown inspection and the lack of training samples for deep learning in engineering sites, and provides key technical support for intelligent control of the stability of a shield excavation face.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring technology for tunnel engineering, and more specifically, to a ground-penetrating radar detection system and method for slurry balance shield tunnels. Background Technology

[0002] Faced with the complex and ever-changing geological conditions during the construction of cross-river and cross-sea tunnels, and the trend of shield tunnels developing towards high water pressure, large diameter, and long distances, slurry-balanced shield tunnels have been widely used in modern tunnel engineering due to their advantages of stable and easily adjustable support pressure. During slurry-balanced shield tunneling, pressurized slurry acts on the excavation face and permeates into the strata. Bentonite particles in the slurry gradually deposit in the soil pores, forming a watertight or slightly permeable mud film on the excavation face. This is crucial for establishing slurry pressure balance and maintaining the stability of the excavation face. The formation state and thickness of the mud film directly affect filtration loss control, support pressure transmission, and the overall stability of the excavation face.

[0003] However, in existing engineering practices, the assessment of mud film condition largely relies on indirect inference from mud parameters and tunneling parameters, or on manual observation through shutdown and hatch opening. These methods not only struggle to achieve in-situ, real-time monitoring during construction but also suffer from latency, and shutdown and hatch opening inspections carry high risks and costs. Although ground-penetrating radar (GPR) detection technology offers advantages such as speed, non-destructive operation, and minimal disruption to construction, the complex tunnel environment, with its uneven excavation face and various mechanical and electromagnetic interferences, makes radar signal processing and interpretation extremely difficult. Currently, radar image interpretation still heavily relies on human experience, making misjudgments highly likely under severe interference conditions. Therefore, a ground-penetrating radar detection system and mud film identification method for slurry balance shield tunneling are urgently needed. Summary of the Invention

[0004] In view of the above-mentioned problems existing in the prior art, this application provides a ground-penetrating radar detection system and method for slurry balance shield tunneling.

[0005] According to a first aspect of this application, a ground-penetrating radar detection method for slurry balance shield tunneling is provided, comprising: S1: injecting Gaussian white noise and power frequency interference into a geological numerical model to obtain a noisy simulated radar image, and extracting corresponding physical attribute labels from the geological numerical model to obtain a training dataset containing pairs of noisy simulated radar images and physical attribute labels; S2: performing supervised training on a deep learning network with a multi-branch decoding structure based on the training dataset to obtain a pre-trained inversion model; S3: acquiring real-time echo signals and cutterhead position data during shield tunneling, and performing spatial correction and normalization processing on the real-time echo signals based on the cutterhead position data to obtain a standardized real-time image; S4: inputting the standardized real-time image into the pre-trained inversion model for feature extraction and multi-task decoding inference to obtain a stratigraphic structure probability map reflecting the distribution of strata categories ahead of the excavation face; S5: performing a maximum probability index search on the stratigraphic structure probability map to obtain a stratigraphic structure label map; S6: performing connectivity identification and pixel height statistics on mud film category regions based on the stratigraphic structure label map to obtain the mud film thickness.

[0006] According to a second aspect of this application, a ground-penetrating radar detection system for a slurry balance shield tunneling machine is provided, comprising: a shield cutterhead, a recessed mounting base, a radar transceiver module, a protective encapsulation assembly, a signal transmission cable, a central slip ring assembly, and a data processing terminal; wherein, the shield cutterhead includes a frame composed of a plurality of radially arranged support spokes and a closed panel interlaced and welded or bolted between adjacent support spokes; the recessed mounting base is integrally formed or welded to a predetermined position on the closed panel; the radar transceiver module is embedded in the internal space of the recessed mounting base, and the detection beam emitting surface of the radar transceiver module faces the excavation face; the protective encapsulation assembly covers the side of the radar transceiver module facing the excavation face; wherein, one end of the signal transmission cable is electrically connected to the data interface of the radar transceiver module, and the signal transmission cable extends to the center of the shield cutterhead and connects to the rotating end of the central slip ring assembly; the stationary end of the central slip ring assembly is electrically connected to the data processing terminal; wherein, the data processing terminal is used to execute the ground-penetrating radar detection method for a slurry balance shield tunneling machine as described in the first aspect.

[0007] Compared with existing technologies, the ground-penetrating radar detection system and method for slurry balance shield tunneling provided in this application significantly improves the anti-interference ability and recognition robustness of the deep learning model under complex dynamic conditions of shield tunneling by introducing environmental noise during the training phase, eliminating misjudgments caused by cutterhead vibration or electromagnetic interference. Furthermore, it enables dynamic and non-destructive monitoring throughout the entire construction process, allowing real-time acquisition of mud film thickness information without stopping the machine or opening the hatch, greatly reducing construction risks and improving operational efficiency. By converting complex radar signals into intuitive geological structure labels and quantified thickness values, this application provides a scientific basis for optimizing mud slurry ratios and dynamically adjusting support pressure, ensuring the safety of shield tunneling under high water pressure and complex geological conditions. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a schematic diagram of the relative permittivity according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of a simulated radar signal image according to an embodiment of this application.

[0011] Figure 3 This is a schematic diagram of a layered geological structure according to an embodiment of this application.

[0012] Figure 4 This is a flowchart of a ground-penetrating radar detection method for a slurry balance shield tunnel according to an embodiment of this application.

[0013] Figure 5 This is a schematic diagram of a radar transceiver module installed on the cutterhead of a tunnel boring machine according to an embodiment of this application.

[0014] Figure 6 A side view of a radar transceiver module mounted on a tunnel boring machine cutterhead according to an embodiment of this application.

[0015] Figure 7 This is a schematic diagram of the method for fixing the radar transceiver module on the shield cutterhead according to an embodiment of this application.

[0016] In the diagram: 1. Tunnel boring machine cutterhead; 2. Spokes; 3. Enclosed panel; 4. Radar transceiver module; 5. Bolts; 6. Protective encapsulation components; 7. Signal transmission cables. Detailed Implementation

[0017] The embodiments of this application will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0018] To better understand the technical solution of this application, the key data images involved in this application will first be explained:

[0019] like Figure 1 As shown in the figure, this diagram illustrates the relative permittivity image output by the deep learning model according to an embodiment of this application. The image is rendered in grayscale or pseudo-color, where the shades of color represent different predicted permittivity values, visually reflecting the differences and distribution of electromagnetic properties of the strata (including mud film, seepage zone, and undisturbed soil) in front of the excavation face. In the figure, the high permittivity region corresponds to the formed mud film.

[0020] like Figure 2 As shown in the figure, this diagram illustrates a simulated radar signal image according to an embodiment of this application. This image serves as input or training samples for the model; its horizontal axis represents the trace number, and its vertical axis represents the two-way travel time of electromagnetic waves, which can be converted into depth. The variations in depth presented in the figure represent the amplitude intensity of the received signal, simulating the propagation and reflection of electromagnetic waves in a pre-defined "mud film-permeable zone-undisturbed soil" layered geological model.

[0021] like Figure 3 As shown, this figure illustrates a layered stratigraphic structure image according to an embodiment of this application. The image is rendered in grayscale and overlaid with layer boundary lines, where different grayscale areas correspond to different stratigraphic types: the uppermost thin layer is a mud film, the middle layer is a permeable zone, and the lowermost layer is undisturbed soil; the three are distributed sequentially. This structural diagram is used to visually demonstrate the spatial distribution relationship of the preset stratigraphic layers and serves as the geometric and stratification basis for subsequent radar signal numerical simulation and deep learning model training / validation.

[0022] This application proposes a ground-penetrating radar detection method for slurry balance shield tunneling machines. Figure 4 The flowchart below shows a ground-penetrating radar detection method for a slurry balance shield tunnel according to an embodiment of this application. Figure 4 As shown, the ground-penetrating radar detection method for slurry balance shield tunneling according to an embodiment of this application includes:

[0023] S1: Gaussian white noise and power frequency interference are injected into the geological numerical model to obtain noisy simulated radar images. Corresponding physical attribute labels are then extracted from the geological numerical model to obtain a training dataset containing paired noisy simulated radar images and physical attribute labels. It is understandable that, given the extreme difficulty in obtaining a large number of accurately labeled mud film radar echo samples at the slurry balance shield tunneling site, generating a large amount of diverse labeled data through numerical simulation can effectively solve the problem of insufficient training samples caused by missing labels at the site. Simultaneously, due to the uneven excavation surface and various mechanical and electromagnetic interferences at the tunnel site, actively injecting Gaussian white noise and power frequency interference that match the characteristics of environmental noise into the simulated signal can simulate a complex underground electromagnetic detection environment. This significantly enhances the anti-interference ability and recognition robustness of the deep learning model under the complex dynamic conditions of the shield tunneling, eliminating misjudgments caused by environmental interference.

[0024] Specifically, the geological numerical model refers to a geoelectric model matrix containing mud film layers, permeable zones, and undisturbed soil layers, constructed based on preset geological parameters and radar system parameters. The physical attribute labels include pixel-level true labels for dielectric constants and stratigraphic structures, with the stratigraphic structure labels using numerical gradients to divide the spatial grid into discrete semantic regions. The noisy simulated radar image refers to a two-dimensional profile image formed by superimposing noise signals simulating electromagnetic interference characteristics from the simulated field onto the clean electromagnetic echo generated by forward modeling.

[0025] In one specific implementation of this application, step S1 includes: performing spatial grid mapping and randomization on preset geological parameters and radar system parameters to obtain a geoelectric model matrix containing the spatial dielectric constant gradient distribution; performing time-domain finite-difference forward modeling and injecting environmental noise on the geoelectric model matrix to obtain a noisy simulated radar image; extracting physical attribute labels from the geoelectric model matrix to obtain physical attribute labels and stratigraphic structure labels corresponding to the pixel level of the noisy simulated radar image; and encapsulating the noisy simulated radar image with the physical attribute labels and stratigraphic structure labels to obtain a training dataset.

[0026] Specifically, spatial grid mapping and randomization are performed on the preset geological parameters and radar system parameters. Based on the spatial grid step size in the radar system parameters, a two-dimensional Cartesian coordinate grid system is initialized, and the dielectric constant range of the undisturbed soil layer is determined according to the detailed engineering geological survey report. The Monte Carlo random sampling method is used to generate specific mud film thickness values ​​and formation interface undulation functions, as well as the relative dielectric constant distribution along the depth direction in the geoelectric model. Using piecewise function definition:

[0027]

[0028] In the above formula, For depth The relative permittivity distribution function at that location, For the formation depth coordinate variable, The relative permittivity of the mud film layer is determined based on the bentonite particle content and moisture content in the shield tunneling mud mix. For example, in simulations of thick mud films with high bentonite content, it can be set to a random value between 30 and 60. This is the background relative permittivity of the undisturbed soil layer, which is usually determined based on the detailed engineering geological survey report. For example, it can be set to 6 to 12 for silty clay layers. The thickness limit of the mud film layer is set within a random range of 0mm to 15mm according to the requirements of the tunnel boring machine (TBM) process. The depth limit of the infiltration zone is typically taken as 2 to 4 times the thickness of the mud film. The attenuation coefficient of the dielectric constant of the permeable zone represents the permeability gradient of the mud intruding into the formation, and its value is positively correlated with the soil porosity. This is a random disturbance term within the undisturbed soil, used to simulate geological heterogeneity. Its value is usually set to within ±5% of the mean dielectric constant of the undisturbed soil.

[0029] Subsequently, a finite-difference time-domain (FDTD) forward modeling simulation and environmental noise injection were performed on the geoelectric model matrix. Using the FDTD algorithm, virtual transmitting and receiving antennas were configured, and a Ricker wavelet was loaded as the excitation source onto the geoelectric model matrix. Maxwell's curl equation was solved to simulate the reflection and refraction of electromagnetic waves at the interfaces between mud film, permeable zone, and undisturbed soil layers. The time-domain electric field intensity changes at the receiving antenna were recorded to generate a clean synthetic radar echo. To simulate real-world conditions, random white noise conforming to a Gaussian distribution and power frequency interference noise were generated and superimposed on the clean echo to generate a noisy simulated radar image. The noisy signal generation model employed additive noise superposition.

[0030]

[0031] In the formula, For the generated first The Way is The amplitude of the noisy analog radar image signal at any given time. The amplitude of the noise-free, clean radar echo signal obtained from forward modeling. For two-way travel time variables, The horizontal position variable along the survey line, The noise amplitude adjustment coefficient is determined by a preset signal-to-noise ratio (SNR), and is usually determined based on the noise floor of the measured radar data of the tunnel. For example, when the SNR is set between 10 dB and 30 dB, The intensity is then dynamically adjusted to satisfy the intensity mapping. The environmental interference weighting factor is a weighting factor for the specific power frequency interference intensity of the tunnel boring machine (TBM). For example, when the main drive motor of the TBM has a high power, resulting in significant electromagnetic interference, the weighting factor is increased. Values ​​(such as 1.5 to 2.5) are used to simulate radar response under extreme noise conditions. It is a sequence of random variables that follow a standard normal distribution.

[0032] Finally, physical attribute labels were extracted and the dataset was encapsulated. The geoelectric model matrix was analyzed, and the relative permittivity values ​​of each pixel were extracted to generate normalized physical attribute labels. Simultaneously, based on the permittivity gradient, mud film regions were labeled as 0, permeable zones as 1, and undisturbed soil as 2, generating stratigraphic structure labels. The noisy simulated radar images were then paired and encapsulated with the corresponding physical attribute labels and stratigraphic structure labels at the pixel level, ultimately constructing the training dataset.

[0033] In one specific embodiment of this application, the thickness of the mud film, the thickness of the permeable zone, the type of undisturbed soil, and the interface undulations are systematically varied to generate thousands of different geological cases. The mud film is set as a relatively high dielectric constant thin layer, with a simulated thickness ranging from 0 mm to 15 mm. The dielectric constant of the permeable zone is set to decrease exponentially from the mud film value to the undisturbed soil value to simulate the intrusion gradient of mud particles. Forward modeling is performed on each case using the same device parameters as actual radar, and the program synchronously outputs the true dielectric constant profile and stratigraphic category label map for that scenario. By adding Gaussian white noise and power frequency interference of different intensities to the synthetic data, and uniformly scaling all simulated radar images to a fixed size, the training dataset is constructed, thereby providing diverse sample support for the supervised training of subsequent deep learning inversion models.

[0034] S2: Supervised training of a deep learning network with a multi-branch decoding structure is performed on the training dataset to obtain a pre-trained inversion model. It is understandable that traditional ground-penetrating radar signal interpretation heavily relies on human experience, making it prone to misjudgment in the complex interference environment of tunnel sites. The pre-trained inversion model obtained through supervised training enables automatic and intelligent interpretation of radar echo signals, transforming complex radar waveform data into intuitive, physically meaningful images of geological attributes. Simultaneously, the multi-branch decoding structure allows the model to simultaneously handle dielectric constant regression and geological structure classification tasks while sharing feature extraction capabilities, thereby improving the model's accuracy and comprehensiveness in identifying subtle geological features such as mud films, permeable zones, and undisturbed soil.

[0035] In one specific implementation of this application, step S2 includes: randomly initializing the weight parameters of the initial deep learning network, the initial deep learning network including a shared feature extractor and parallel regression and classification branches; inputting noisy simulated radar images from the training dataset into the initial deep learning network for forward propagation calculation, and calculating the multi-task joint total loss value including regression mean square error and classification cross-entropy by combining the corresponding physical attribute labels and stratigraphic structure labels; and iteratively updating the weight parameters of the initial deep learning network multiple times based on the joint total loss value until the network converges to obtain the pre-trained inversion model.

[0036] Specifically, the first step is to construct and initialize a multi-branch encoder-decoder network architecture. The data dimensions in the training dataset are analyzed to construct a shared feature extractor. This encoder, composed of stacked convolutional layers, batch normalization layers, and max-pooling layers, is used to extract high-dimensional latent features from the input image. Two independent decoder branches are connected in parallel at the end of the encoder. The regression branch uses deconvolutional layers to progressively recover the spatial dimensions and employs the ReLU activation function at its end to generate a continuous-valued dielectric constant map. The classification branch has a similar structure to the regression branch, but its end corresponds to the number of classes in the stratigraphic structure label, and it outputs a probability map using the Softmax activation function. He initialization or Xavier initialization methods are used to randomly assign values ​​to the network's convolutional kernel weights and biases to generate an initial deep learning network with a complete topology but whose parameters have not yet converged.

[0037] Subsequently, dual-stream forward propagation and multi-task joint loss calculation are performed. Noisy simulated radar images are batch-sampled from the training dataset as input, and corresponding physical attribute labels and stratigraphic structure labels are extracted as ground truth. The noisy simulated radar images are fed into the initial deep learning network. After being split by the encoder, the regression branch outputs the predicted dielectric constant, and the classification branch outputs the predicted structure distribution. The loss values ​​for the two tasks are calculated separately: the regression task calculates the mean squared error between the predicted value and the ground truth, and the classification task calculates the cross-entropy loss between the predicted probability distribution and the ground truth label. A dynamic weighting factor is introduced to weight and sum the two loss values ​​to obtain the joint total loss value used to guide optimization. The multi-task joint loss function is composed of a weighted regression loss term and a classification loss term, and its calculation formula is as follows:

[0038]

[0039] In this formula, This represents the total combined loss value for the current batch of multi-task operations. These are the loss weighting coefficients for the regression task; For the loss weight coefficients of the classification task; This represents the total number of pixels in the current training batch. The output of the regression branch of the network The predicted relative permittivity value of each pixel; For the first The true value of the physical attribute label corresponding to each pixel; This represents the total number of categories of stratigraphic structures. For the first Each pixel belongs to the category The truth value of; The output of the network classification branch Each pixel belongs to the category The predicted probability.

[0040] Finally, gradient backpropagation and adaptive parameter optimization are performed. Based on the joint total loss, the gradient vector of each layer parameter in the network is calculated using the chain rule. The Adam adaptive moment estimation algorithm is employed, adaptively adjusting the learning rate of each parameter based on the first and second moment estimates of the gradient, and updating the network parameters for the current batch using the calculated update step size. The formula for this parameter update rule is as follows:

[0041]

[0042] In this formula, For the updated network parameters; The network parameters at the current moment before the update; The global learning rate; Correct the first-order moment estimate for the gradient bias; Correct the second-order moment estimate for the gradient bias; To prevent smoothing terms with a denominator of zero, the aforementioned forward and backward propagation processes are repeated until the joint total loss value no longer decreases significantly on the validation set or reaches the preset number of iterations. The final parameters are then saved, and the pre-trained inversion model is output.

[0043] In one specific embodiment of this application, the constructed deep learning network takes a single-channel radar image as input and outputs two branches. One branch outputs a single-channel dielectric constant regression map, where each pixel value represents the predicted relative dielectric constant; the other branch outputs a three-channel stratigraphic classification probability map, where the values ​​of each pixel in the three channels represent the probability that it belongs to mud film, permeable zone, or undisturbed soil, respectively. During training, thousands of labeled samples generated by the geoelectric model are used, and GPU-accelerated computation is employed to continuously minimize the pixel-level difference between the predicted image and the actual physical label, thereby enabling the model to accurately extract the geometric boundaries and electrical characteristics of the mud film layer from complex, noisy simulated radar images.

[0044] S3: Acquire real-time echo signals and cutterhead position data during the shield tunneling process, and perform spatial correction and normalization processing on the real-time echo signals based on the cutterhead position data to obtain standardized real-time images. It is understandable that shield tunneling conditions are extremely complex. Affected by uneven ground hardness, mud cake friction, or drive system response delays, the cutterhead rotation is not uniform but accompanied by instantaneous angular velocity fluctuations (i.e., stick-slip phenomenon) and abrupt changes in angular acceleration. This dynamic characteristic leads to the Doppler effect and integral ambiguity. That is, when the cutterhead angular velocity is too high, the physical integration area of ​​the radar single-channel signal increases, resulting in a decrease in lateral resolution; and when instantaneous jamming or abrupt changes in angular acceleration occur, the coupling state between the radar and the detection interface will oscillate, introducing non-geological mechanical noise. If only the standard angle interpolation method is used, low signal-to-noise ratio anomalous data and high-fidelity data will participate in image reconstruction with the same weight, directly causing artifacts in the final polar coordinate radar matrix, severely interfering with subsequent identification of small mud films. Furthermore, deep learning models require input data to have consistent size and amplitude range. Therefore, spatial correction and normalization must be performed to transform the original, dynamically disturbed signals into standardized real-time images. This ensures that every pixel in the reconstructed image primarily originates from high-fidelity radar echoes, providing a high-quality data foundation for subsequent intelligent recognition of small targets.

[0045] In one specific implementation of this application, step S3 includes: generating an original polar coordinate radar matrix based on real-time echo signals and cutter head position data; transforming the original polar coordinate radar matrix from the polar coordinate system to the Cartesian rectangular coordinate system to obtain a reconstructed Cartesian domain image; and performing signal standardization enhancement and tensor adaptation on the reconstructed Cartesian domain image to obtain a standardized real-time image.

[0046] Specifically, firstly, a raw polar coordinate radar matrix is ​​generated based on real-time echo signals and cutterhead position data. By associating real-time echo signals with cutterhead position data (such as angular coordinates), the single-channel echoes collected by the radar along the cutterhead rotation path can be mapped to the polar coordinate system, thereby synthesizing a two-dimensional radar image reflecting the annular scanning section in front of the excavation face.

[0047] In one specific implementation of this application, a raw polar coordinate radar matrix is ​​generated based on real-time echo signals and cutter head position data. This includes mapping the radar data to a polar coordinate grid using timestamp alignment and standard angle interpolation. This process utilizes timestamp alignment technology to bind each real-time echo signal to an accurate angle in the cutter head position data. Since the acquired signals may be unevenly distributed in terms of angle, an interpolation algorithm is used to map the signals onto a uniformly spaced angle grid, constructing the raw polar coordinate radar matrix. Each column of the matrix represents a specific angular position, and each row represents the two-way travel time depth of the electromagnetic wave.

[0048] However, research revealed that the aforementioned implementation methods neglected the nonlinear coupling relationship between the shield cutterhead rotation dynamics and the radar echo signal fidelity. In slurry shield tunneling, due to uneven ground hardness, mud cake friction, or drive system response delays, the cutterhead rotation is not uniform but accompanied by instantaneous angular velocity fluctuations (i.e., stick-slip phenomenon) and abrupt changes in angular acceleration. This dynamic characteristic leads to the Doppler effect and integral ambiguity. Specifically, when the cutterhead angular velocity is too high, the physical integration area of ​​the radar single-channel signal increases, resulting in a decrease in lateral resolution. Furthermore, when instantaneous pauses or abrupt changes in angular acceleration occur, the coupling state between the radar and the detection interface oscillates, introducing non-geological mechanical noise. The angle interpolation method in the aforementioned implementation treats all acquired scan channels as data points of equal quality for geometric filling. This means that low signal-to-noise ratio anomalous data acquired during severe cutterhead vibration or overspeed rotation participates in image reconstruction with the same weight as high-fidelity data acquired during stable rotation, directly causing artifacts in the final polar coordinate radar matrix and severely interfering with subsequent identification of minute mud films.

[0049] Based on this, an improved mechanism is proposed, which aims to construct a data reconstruction method based on dynamic confidence weighting. This method achieves adaptive denoising and high-fidelity reconstruction of radar data by quantifying the influence of the cutterhead motion state on the echo quality. Specifically, in another implementation of this application, an original polar coordinate radar matrix is ​​generated based on the real-time echo signal and cutterhead position data. This includes: performing echo kinematic state calculation on the real-time echo signal and cutterhead position data to obtain a motion state vector sequence; determining the confidence weight vector based on the motion state vector sequence; and performing grid reconstruction on the real-time echo signal and cutterhead position data based on the confidence weight vector to obtain the original polar coordinate radar matrix.

[0050] First, to accurately capture the minute vibrations and instantaneous velocity changes of the tunnel boring machine cutterhead during excavation, echo kinematic state calculation is required. This process abandons the simple approach of directly differentiating the raw angle data, instead utilizing a Gaussian smoothing differential operator under a sliding window to convolve the timestamp of each frame of real-time echo signal with the cutterhead rotation angle. This operation effectively suppresses high-frequency noise caused by sensor quantization errors, thereby accurately calculating the instantaneous angular velocity and angular acceleration at each signal acquisition moment, achieving a quantitative description of the cutterhead rotation dynamics. The calculation formula for the echo kinematic state calculation is as follows:

[0051]

[0052] in, The instantaneous angular velocity at the time of signal acquisition in the k-th frame is used to measure the speed of rotation. The instantaneous angular acceleration during the acquisition of the k-th frame signal is used to characterize the vibration intensity; This represents a Gaussian differential convolution kernel, used to extract smooth velocity change trends; This represents the rotation angle of the cutter head.

[0053] Subsequently, based on the kinematic parameters obtained from the above calculations, a signal fidelity evaluation index is constructed. This index is based on a physical assumption: when the cutterhead is at a stable speed close to its rated speed determined by the radar pulse repetition frequency and lateral sampling rate, the coupling state between the radar and the medium is optimal, and the data reliability is highest; conversely, when the speed deviates from the optimal value or there is severe acceleration or deceleration (such as vibration or impact), the data reliability decreases exponentially. Therefore, by introducing a speed deviation penalty term and a vibration stability penalty term, the reliability weight of each echo signal is calculated to obtain the reliability weight vector. The formula for calculating the reliability weight is:

[0054]

[0055] In the formula, The confidence weight represents the k-th signal; The optimal observation angular velocity is determined based on the pulse repetition frequency of the ground-penetrating radar and the preset lateral sampling resolution. For example, when the radar pulse frequency is 1000Hz and a signal is required to be collected every 0.1 degrees, Set to 100 degrees / second; This is the speed tolerance coefficient, used to control the sensitivity of the evaluation mechanism to speed fluctuations; This is the mechanical vibration sensitivity factor. The larger the value, the heavier the penalty for sudden changes in angular acceleration. Its value can be determined based on the historical vibration statistics of the tunnel boring machine when excavating in specific strata. For example, in high-vibration conditions such as rock strata, a larger value is set. Values ​​(e.g., 1.5 to 3.0) are used to enhance the suppression of abrupt changes in angular acceleration data. This process quantifies the physical validity of each signal, automatically identifying and labeling anomalous data that is contaminated by vibration or has reduced resolution.

[0056] Finally, grid reconstruction based on Nadalaya-Watson kernel estimation is performed. When constructing the original polar coordinate radar matrix, the traditional linear interpolation method is abandoned, and a kernel regression estimation strategy is adopted. For each node on the uniform angle grid of the target, weighted reconstruction is performed using real signals acquired from the surrounding area. The weights here not only consider the spatial distance between the signal acquisition location and the target grid point, but also deeply integrate the previously calculated dynamic confidence weights. The grid reconstruction calculation is as follows:

[0057]

[0058] In the formula, For the reconstructed first Each angle grid at depth Polar coordinate magnitude at; It is the amplitude of the original acquired signal; Represents a spatial smoothing kernel function (e.g., a Gaussian kernel). This is a bandwidth parameter, whose value is preset based on the angular sampling density of the radar scan to balance the resolution of the reconstructed image and the noise suppression effect. and These represent the set angle of the target grid and the actual sampling angle of the original signal, respectively. This step achieves the removal of spurious data and the retention of true polar coordinate data through a weighted mechanism that prioritizes high-quality data and filters out low-quality data.

[0059] This specific implementation effectively addresses the shortcomings of traditional interpolation methods in handling the complex dynamic characteristics of tunnel boring machine (TBM) construction by introducing a dynamic reliability weighting mechanism. By establishing a data reconstruction process that adapts to the cutterhead stick-slip effect and vibration interference, it ensures that every pixel in the reconstructed image primarily originates from high-fidelity radar echoes. This significantly improves the signal-to-noise ratio and lateral resolution of polar coordinate radar images, particularly by effectively removing non-geological noise artifacts caused by cutterhead jamming or overspeeding. This allows for the clear representation of mud film layers with micron-level precision in the reconstructed images, providing a high-quality data foundation for subsequent intelligent identification of small targets.

[0060] After obtaining the original polar coordinate radar matrix, Cartesian coordinate resampling and spatial reconstruction are performed to obtain the reconstructed Cartesian domain image. A Cartesian coordinate grid for the target image is defined. The bilinear interpolation algorithm is used to map polar coordinate grid points to rectangular coordinate grid points, completing the spatial transformation from the annular scan domain to the linearly expanded domain. The calculation formula is as follows:

[0061] in ,

[0062] In the formula, These are the pixel values ​​in the reconstructed Cartesian coordinate system. Radar installation radius; denoted as , where is the average propagation speed of electromagnetic waves in the medium.

[0063] Finally, the reconstructed Cartesian domain image is normalized and enhanced using tensor fitting. This involves sequentially performing DC offset removal, frequency domain filtering, background clutter suppression (e.g., subtracting the sliding window average), and gain compensation. The processed image amplitude is then normalized using a linear mapping. The interval is scaled to the same dimension as the input layer of the trained model, and its normalization formula is:

[0064]

[0065] In the formula, To standardize the pixel values ​​of real-time images; This is the preprocessed intermediate image data. This represents the pixel value of the preprocessed intermediate image data.

[0066] In one specific embodiment of this application, the ground-penetrating radar on the tunnel boring machine cutterhead automatically collects data at set time intervals, and the acquisition host synchronously records the cutterhead rotation angle, advance distance, and time. Laser calibration is used to ensure that the antenna transmitting surface is perpendicular to the excavation face. Non-geological noise artifacts caused by cutterhead jamming or overspeed are filtered out through a dynamic confidence weighting mechanism. The acquired single-channel echo is correlated with the rotation angle coordinates, and after DC removal, filtering, background removal, and resampling operations, a standardized 256×256 pixel two-dimensional radar image is finally generated, so that the mud film layer with micron-level precision can be clearly presented in the image.

[0067] S4: Standardized real-time images are input into a pre-trained inversion model for feature extraction and multi-task decoding inference to obtain a stratigraphic structure probability map reflecting the distribution of strata types ahead of the excavation face. It is understandable that the environment at the shield tunnel site is extremely complex. Affected by uneven excavation faces and various mechanical and electromagnetic interferences, the original radar echo signals are difficult to directly and accurately identify manually, and traditional interpretation methods heavily rely on human experience, easily leading to misjudgments. By inputting pre-processed standardized images into the pre-trained inversion model, the powerful nonlinear mapping capabilities of deep learning models can be utilized to automatically and in-situ extract key geological features from complex radar signals and transform them into intuitive, physically meaningful stratigraphic structure probability distributions. This provides scientific data support for the quantitative statistics of mud film thickness, solving the problems of scarce training samples and low recognition accuracy of deep learning in complex environments.

[0068] In one specific implementation of this application, step S4 includes: inputting a standardized real-time image into the shared encoder of a pre-trained inversion model for convolution and downsampling to extract a latent feature tensor containing high-dimensional geological features; inputting the latent feature tensor into the regression decoder branch of the pre-trained inversion model for deconvolution upsampling and linear activation mapping to obtain a predicted dielectric constant map; and inputting the latent feature tensor into the classification decoder branch of the pre-trained inversion model for multi-scale feature fusion and probability normalization calculation to obtain a stratigraphic structure probability map.

[0069] Specifically, deep feature encoding and latent space mapping are first performed, and the generated standardized real-time image is fed into the shared encoder of the pre-trained inversion model. The data flows through a series of convolutional layers inside the encoder for local texture feature extraction, parameter distribution is standardized through batch normalization layers, and nonlinearity is introduced using activation functions. After downsampling and pooling operations, the data outputs a high-dimensional compressed latent feature tensor at the encoder bottleneck layer, thereby achieving the abstract extraction of key geological information from the input radar image.

[0070] Subsequently, physical parameter regression decoding and stratigraphic structure classification decoding are performed in parallel. In the regression decoding stage, the latent feature tensor is imported into the model's regression decoder branch. Through a series of transposed convolutions or upsampling operations, the spatial resolution of the feature map is progressively restored to the original input size. Skip connections are used to fuse shallow detail features from the encoder to enhance boundary sharpness. A linear activation function is applied to the output layer to map the feature values ​​to continuous physical parameter values, generating a predicted dielectric constant map. The calculation formulas for the layer-by-layer decoding and final mapping process of the regression branch are as follows:

[0071]

[0072] In the above formula, The final output is a predicted dielectric constant plot in coordinates. The predicted value at that location; This represents the nonlinear mapping function of the regression decoder network; The input is the latent feature tensor; The linear activation function of the output layer; This is the weight matrix of the final convolutional layer of the regression decoder; This is the convolution operator; This is an upsampling or deconvolution operation used to restore spatial dimensions; This is the bias term for the final convolutional layer of the regression decoder.

[0073] During the classification and decoding stage, the latent feature tensor is simultaneously imported into the model's classification and decoder branch. After restoring spatial resolution through upsampling and multi-scale feature fusion, the number of channels in the network's final output layer is set to equal the number of formation categories (e.g., 3 categories: mud film, permeable zone, and undisturbed soil). The Softmax function is applied to the output feature map along the channel dimension to transform the values ​​into a normalized probability distribution, generating a formation structure probability map. The probability generation process of the classification branch and the Softmax normalization calculation formula are as follows:

[0074]

[0075] In the above formula, For the probability map of stratigraphic structure at location Belongs to the first The probability of a geological formation; For the classification decoder before the Softmax layer The raw logarithmic odds of the channel output; In this embodiment, the total number of stratigraphic categories is [number]. ; For the category index variable used in the summation; It is a natural exponential function.

[0076] In one specific embodiment of this application, real-time radar signals acquired during tunnel boring are preprocessed to form a standardized real-time image input model. The model encoder extracts deep geological features and feeds them to the decoder. The regression branch calculates the predicted relative permittivity value for each pixel, generating an image that reflects the differences in the electromagnetic properties of the strata; regions with high permittivity correspond to mud film. The classification branch outputs a stratigraphic structure probability map reflecting the distribution of strata categories ahead of the excavation face. This probability map uses three channels to represent the probability that each pixel belongs to mud film, permeable zone, and undisturbed soil, respectively. This provides a high-precision probability field basis for subsequently generating a discretized stratigraphic structure label map through maximum probability index search and ultimately calculating the mud film thickness.

[0077] S5: Perform a maximum probability index search on the stratigraphic structure probability map to obtain the stratigraphic structure label map. It should be understood that the stratigraphic structure probability map output by the classification decoding branch of a deep learning model typically contains multiple channels, each reflecting the confidence level of a pixel belonging to a specific stratigraphic category. Since these probability values ​​are continuous floating-point numbers, they cannot be directly used to define stratigraphic boundaries or calculate geometric thickness. By performing a maximum probability index search, the confidence distribution of soft classification can be transformed into a category index for hard classification, thereby clarifying the attribute attribution of each spatial sampling point within the stratigraphic structure. Furthermore, considering the fluctuations in model predictions caused by environmental noise during tunnel boring, introducing spatial morphological smoothing after the index search can effectively eliminate isolated misclassified pixels (such as salt-and-pepper noise), ensuring that the generated label map has physical coherence, thereby improving the accuracy and robustness of mud film recognition.

[0078] In one specific implementation of this application, step S5 includes: performing pixel-by-pixel maximum a posteriori probability decision on the stratigraphic structure probability map to obtain an initial classification matrix; and performing spatial morphological filtering and semantic smoothing on the initial classification matrix to obtain a stratigraphic structure label map.

[0079] Specifically, the first step is to make a pixel-by-pixel maximum a posteriori probability decision. This involves traversing every spatial coordinate point in the stratigraphic structure probability map. The probability vector of a point is extracted along the depth direction, i.e., the channel dimension. The probability values ​​of the pixel in the three channels are compared, and the category index corresponding to the maximum probability is selected as the final category of the pixel. Through this hard classification decision based on the maximum a posteriori probability, an initial classification matrix representing the preliminary distribution of stratigraphic categories is generated. The calculation formula is as follows:

[0080]

[0081] In the above formula, The initial classification matrix output in coordinates The original category index value at the location; A function to find the maximum value of an independent variable, used to return the channel index that maximizes the probability value; For category channel index variables; The total number of categories in stratigraphic classification; The first digit of the input stratigraphic probability map The probability value of the channel at the current pixel position.

[0082] After obtaining the initial classification matrix, spatial morphological filtering and semantic smoothing are then performed. For isolated misclassified pixels that may exist in the initial classification matrix, a [missing information - likely a function or method] is defined. or A sliding window is used as a structuring element to slide across the matrix. For each window coverage area, the mode of pixel categories within the area is counted, or a majority vote is performed. The category of the central pixel of the window is replaced with the category value that appears most frequently in that neighborhood. This smoothing filtering operation based on local neighborhood majority voting removes spatial discontinuities while preserving the continuous features of the mud film, thus obtaining the final stratigraphic structure label map. The calculation formula is as follows:

[0083]

[0084] In the above formula, The final output stratigraphic structure label map is in coordinates Smoothed category value at the location; This is the mode function, used to return the most frequent value in a set; In pixels A local neighborhood set centered on; This represents the original category value at each coordinate point within the neighborhood.

[0085] In one specific embodiment of this application, the probability map output by the multi-task model is post-processed, and the category index (0, 1, or 2) corresponding to the maximum probability in the three channels of each pixel position is taken. By performing this hard classification judgment on all pixels in the image, an initial stratigraphic distribution matrix is ​​generated. Subsequently, a preset sliding window is used to semantically smooth the matrix to eliminate sporadic misclassification points inside or at the edges of the mud film area. The final generated label map clearly shows the distribution of the mud film layer at the top layer with discrete gray levels, providing direct data support for subsequent extraction of mud film layer pixels and calculation of the physical thickness of the mud film using pixel-actual scale calibration coefficients.

[0086] S6: Based on the geological structure label map, connectivity identification and pixel height statistics are performed on mud film category regions to obtain mud film thickness. This can be understood as transforming a qualitative geological structure image into a quantitative physical thickness parameter. Since the classification results output by the deep learning model only reflect the category of spatial pixels and cannot directly guide construction decisions, it is necessary to post-process the label map to extract the geometric features of the mud film layer and evaluate its film continuity and uniformity, thereby calculating the accurate physical thickness. This solves the problem of the lag in traditional methods that rely on indirect inference from mud parameters or observation during shutdown, providing an in-situ, real-time, and non-destructive monitoring method for intelligent control of shield tunneling face stability.

[0087] In one specific implementation of this application, step S6 includes: extracting pixel regions belonging to the mud film category from the stratigraphic structure label map and performing connected component analysis to obtain an effective mud film mask; determining the mud film state based on the effective mud film mask; and, in response to the mud film state being a film already formed, performing pixel-level thickness statistics and physical space mapping on the effective mud film mask to obtain the mud film thickness.

[0088] Specifically, firstly, all pixel regions belonging to the mud film category (index value 0) are extracted from the stratigraphic structure label map to generate a binarized candidate mask. Then, connected component analysis is performed on this mask to calculate the span width of the largest connected region in the horizontal direction (i.e., along the cutterhead scanning path). This span width is compared with a preset continuity threshold. If it exceeds the threshold, it is determined that a valid mud film has formed and a valid mud film mask is generated; otherwise, it is determined that no film has formed. The calculation formula for determining the film formation status is as follows:

[0089]

[0090] In the above formula, This represents the mud film formation status of the output, where 1 indicates that a film has formed and 0 indicates that a film has not formed. This is the set of all independent connected components extracted from the binarized mask; For a single connected component object in the set; For connected components The number of pixels representing the projection width in the horizontal direction; The minimum continuous width threshold for determining whether mud film has formed effectively.

[0091] After confirming that the mud film has formed, pixel-level thickness statistics are further performed. By scanning the effective mud film mask column by column, the number of pixels with a value of 1 in each column is counted to obtain the instantaneous pixel height at each location. The pixel heights of all effective columns are summed and their arithmetic mean is calculated to obtain the average pixel height. Finally, this average pixel height is multiplied by a calibration coefficient to obtain the final physical thickness value. The integral formula for calculating the physical thickness of the mud film is as follows:

[0092]

[0093] In the above formula, The final calculated physical thickness of the mud film, in millimeters; The calibration coefficient, i.e., the preset conversion factor between pixel size and physical length, is the longitudinal resolution determined by the ground-penetrating radar's time sampling rate and the propagation speed of electromagnetic waves in the detection medium (such as mud film). It is usually quantized in millimeters per pixel, assuming the ground-penetrating radar's time sampling interval. This allows each pixel to represent a two-way travel time of 0.1 ns, while the wave speed of electromagnetic waves in the mud film medium... Approximately 60 mm / ns. Therefore, calculate the calibration coefficient: The above are merely examples for illustration and do not constitute specific limitations. It is the set of coordinate indices of the effective mud film area in the horizontal direction, that is, the set of columns where mud film exists; Representative set The number of elements, i.e., the total width of the effective mud film; Represents the horizontal coordinate index; Represents the coordinate index of the vertical depth direction; Represents the effective mud mask in coordinates The pixel value at that location.

[0094] In one specific embodiment of this application, after inverting the generated standardized real-time image to obtain a stratigraphic structure probability map, a stratigraphic structure label map is generated through a maximum probability index search. The region corresponding to the excavation face at the top of this label map is analyzed. If a continuously distributed mud film pixel region with a width exceeding a preset threshold exists, it is determined that a valid mud film has been formed. Subsequently, the pixel height occupied by the mud film region in the vertical depth direction is calculated. The pixel height is then converted to physical size using calibration coefficients. The final output is the actual thickness of the mud film. It can be displayed in real time on the monitoring interface and can automatically alarm when the thickness is insufficient according to the preset alarm threshold, thereby guiding construction personnel to optimize the mud ratio or adjust the support pressure.

[0095] In summary, the ground-penetrating radar detection method for slurry balance shield tunneling machines according to the embodiments of this application has been elucidated. It establishes a geoelectric numerical model including mud film, permeable zone, and undisturbed soil, and actively injects Gaussian white noise and power frequency interference to simulate the complex electromagnetic environment of the site. This pre-constructs a large-scale training dataset with precise physical attribute labels, effectively solving the training bottleneck of deep learning in tunnel engineering caused by the difficulty in obtaining labels. Based on this, a multi-branch decoding network architecture is used to simultaneously infer the geological structure and physical parameters, and the real-time signal is spatially corrected by combining shield cutterhead position data, ensuring the fidelity and standardization of the radar echo in the spatial dimension. Finally, morphological connectivity recognition and pixel calibration are used to accurately extract the mud film thickness.

[0096] The following is in conjunction with the appendix Figure 5 To be continued Figure 7 The hardware structure of the ground-penetrating radar detection system for slurry balance shield tunneling machines provided in this application is described in detail. According to an embodiment of this application, the ground-penetrating radar detection system for slurry balance shield tunneling machines includes: a shield cutterhead 1, a recessed mounting base, a radar transceiver module 4, a protective encapsulation assembly 6, a signal transmission cable 7, a central slip ring assembly, and a data processing terminal. Figure 5 This is a schematic diagram illustrating the installation of a radar transceiver module on the cutterhead of a tunnel boring machine according to an embodiment of this application. Figure 5 As shown, the mechanical carrier of the detection system is the shield cutterhead 1, which is mainly used to cut the soil in front and is the main mechanical structure that ensures the shield machine can move forward. The shield cutterhead 1 consists of a frame composed of several supporting spokes 2 arranged radially, and closed panels 3 are welded or bolted between adjacent supporting spokes 2.

[0097] Figure 6 This is a side view of a radar transceiver module mounted on the cutterhead of a tunnel boring machine according to an embodiment of this application. Figure 6As shown, the radar transceiver module 4 is installed at a predetermined position on the enclosed panel 3 (that is, an area on the enclosed panel 3 that avoids the cutting tool and can directly face the excavation face). At this installation position, the plate of the enclosed panel 3 is recessed towards the non-excavation face side (i.e., inward), thereby integrally forming or welding a recessed mounting base, which provides accommodating space for the installation of the radar equipment. The radar transceiver module 4 is embedded in the internal space of the recessed mounting base, and there is a gap between the support spokes 2 and the enclosed panel 3 to allow the cut soil to pass through the cutterhead and enter the slurry chamber. During installation, laser calibration is required to ensure that the antenna transmitting surface of the radar transceiver module 4 faces the excavation face, and its central normal is as perpendicular to the excavation face as possible, thereby ensuring the vertical incidence and reflection effect of electromagnetic waves.

[0098] Figure 7 This is a schematic diagram illustrating the fixing method of the radar transceiver module on the shield cutterhead according to an embodiment of this application. Figure 7 As shown, the radar transceiver module 4 is securely mounted on this recessed mounting base by bolts 5 and a protective encapsulation component 6 (i.e., the protective component). To enable the precision electronic equipment to withstand the harsh environment of high pressure, high humidity, and mud erosion during tunnel boring, the protective encapsulation component 6 covers the side of the radar transceiver module 4 facing the excavation face.

[0099] Specifically, the protective encapsulation component 6 includes a wave-transparent waterproof cover and a ceramic wear-resistant plate. The wave-transparent waterproof cover is preferably made of polytetrafluoroethylene (PTFE), a material that combines high strength, wear resistance, and excellent electromagnetic wave penetration. It surrounds and seals the front face and side walls of the radar transceiver module 4 to achieve waterproof isolation. A ceramic wear-resistant plate is fixed to the outermost surface of the wave-transparent waterproof cover (i.e., the soil-facing surface). This ceramic wear-resistant plate is made of alumina ceramic sheets with a thickness of 15-20 mm to directly resist the cutting and erosion of coarse particles in the mud.

[0100] Regarding electrical connections and signal transmission, one end of the signal transmission cable 7 is electrically connected to the data interface of the radar transceiver module 4 to power the radar and transmit echoes. Most of the signal transmission cable 7 is embedded in the enclosed panel 3 and extends to the center of the shield cutterhead 1, connecting to the rotating end of the central slip ring assembly. The stationary end of the central slip ring assembly is electrically connected to the data processing terminal, thereby transmitting the acquired signals in real time to the data processing unit located in the supporting area behind the shield.

[0101] In actual operation, radar transceiver module 4 acquires real-time echo signals during shield tunneling and transmits them back to the data processing terminal via signal transmission cable 7 and the central slip ring assembly. The data processing terminal inputs the processed, standardized real-time image into a pre-trained inversion model, outputs a stratigraphic structure probability map, and further obtains a stratigraphic structure label map through a maximum probability index search. Finally, the data processing terminal performs connectivity identification and pixel height statistics on mud film category regions based on the stratigraphic structure label map to obtain the mud film thickness. Specifically, the system extracts pixel regions belonging to the mud film category from the stratigraphic structure label map and performs connected component analysis to obtain the effective mud film mask. After determining that the mud film state has been formed, pixel-level thickness statistics are performed on the effective mud film mask. The physical domain mud film thickness value is accurately calculated by multiplying the pixel height occupied by the mud film layer in the image with a preset pixel-actual scale calibration coefficient.

[0102] It should be noted that the ground-penetrating radar detection system for slurry balance shield tunneling in this application embodiment is similar in principle to the aforementioned ground-penetrating radar detection system and method for slurry balance shield tunneling. Therefore, the implementation process, implementation principle, and beneficial effects of the ground-penetrating radar detection system for slurry balance shield tunneling can be found in the description of the implementation process, implementation principle, and beneficial effects of the aforementioned method, and will not be repeated.

[0103] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A ground-penetrating radar detection method for slurry balance shield tunnels, characterized in that, include: S1: Gaussian white noise and power frequency interference are injected into the geological numerical model to obtain noisy simulated radar images, and corresponding physical attribute labels are extracted from the geological numerical model to obtain a training dataset containing pairs of noisy simulated radar images and physical attribute labels. The geological numerical model is a geoelectric model matrix containing mud film layer, permeable zone layer and undisturbed soil layer, constructed based on preset geological parameters and radar system parameters; the physical attribute labels include true labels of dielectric constant and true labels of stratigraphic structure. S2: Supervised training of a deep learning network with a multi-branch decoding structure is performed on the training dataset to obtain a pre-trained inversion model; S3: Acquire real-time echo signals and cutterhead position data during the shield tunneling process, and perform spatial correction and normalization processing on the real-time echo signals based on the cutterhead position data to obtain standardized real-time images. S4: Input the standardized real-time image into the pre-trained inversion model for feature extraction and multi-task decoding inference to obtain a stratigraphic structure probability map reflecting the distribution of strata categories in front of the excavation face; S5: Perform a maximum probability index search on the stratigraphic structure probability map to obtain the stratigraphic structure label map; S6: Based on the stratigraphic structure label map, the connectivity of mud film category regions is identified and pixel height is statistically analyzed to obtain the mud film thickness.

2. The ground-penetrating radar detection method for slurry balance shield tunneling according to claim 1, characterized in that, Step S1 includes: Spatial grid mapping and randomization are performed on the preset geological parameters and radar system parameters to obtain a geoelectric model matrix containing the spatial dielectric constant gradient distribution; Finite-difference forward modeling of the geoelectric model matrix was performed in the time domain, and environmental noise was injected to obtain a noisy simulated radar image. Physical attribute labels are extracted from the geoelectric model matrix to obtain physical attribute labels and stratigraphic structure labels corresponding to the pixel level of the noisy simulated radar image; The noisy simulated radar images are encapsulated with physical attribute labels and stratigraphic structure labels to obtain the training dataset.

3. The ground-penetrating radar detection method for slurry balance shield tunneling according to claim 1, characterized in that, Step S2 includes: The weight parameters of the initial deep learning network are randomly initialized. The initial deep learning network includes a shared feature extractor and parallel regression and classification branches. Noisy simulated radar images from the training dataset are input into the initial deep learning network for forward propagation calculation, and the multi-task joint total loss value, including regression mean square error and classification cross entropy, is calculated by combining the corresponding physical attribute labels and stratigraphic structure labels. The weight parameters of the initial deep learning network are iteratively updated multiple times based on the joint total loss value until the network converges to obtain the pre-trained inversion model.

4. The ground-penetrating radar detection method for slurry balance shield tunneling according to claim 1, characterized in that, Step S3 includes: Based on real-time echo signals and cutter head position data, the original polar coordinate radar matrix is ​​generated; The original polar coordinate radar matrix is ​​transformed from the polar coordinate system to the Cartesian rectangular coordinate system to obtain the reconstructed Cartesian domain image; Signal normalization enhancement and tensor fitting are performed on the reconstructed Cartesian domain image to obtain a normalized real-time image.

5. The ground-penetrating radar detection method for slurry balance shield tunneling according to claim 1, characterized in that, Step S4 includes: Standardized real-time images are input into the shared encoder of the pre-trained inversion model for convolution and downsampling to extract latent feature tensors containing high-dimensional geological features. The latent feature tensor is input into the regression decoder branch of the pre-trained inversion model and deconvolution upsampling and linear activation mapping are performed to obtain the predicted dielectric constant map. The latent feature tensor is input into the classification decoder branch of the pre-trained inversion model to perform multi-scale feature fusion and probability normalization calculations to obtain the stratigraphic structure probability map.

6. The ground-penetrating radar detection method for slurry balance shield tunneling according to claim 1, characterized in that, Step S5 includes: The initial classification matrix is ​​obtained by performing pixel-by-pixel maximum a posteriori probability decision on the stratigraphic structure probability map. Spatial morphological filtering and semantic smoothing are applied to the initial classification matrix to obtain a stratigraphic structure label map.

7. The ground-penetrating radar detection method for slurry balance shield tunneling according to claim 1, characterized in that, Step S6 includes: Extract pixel regions belonging to the mud film category from the stratigraphic structure label map and perform connected component analysis to obtain an effective mud film mask; Determine the state of the mud film based on the effective mud film mask; In response to the mud film state being a film already formed, pixel-level thickness statistics and physical space mapping are performed on the effective mud film mask to obtain the mud film thickness.

8. The ground-penetrating radar detection method for slurry balance shield tunneling according to claim 4, characterized in that, Based on real-time echo signals and cutterhead position data, a raw polar coordinate radar matrix is ​​generated, including: Echo kinematics state calculation is performed on real-time echo signals and cutter head position data to obtain a motion state vector sequence; Determine the confidence weight vector based on the motion state vector sequence; Based on the confidence weight vector, the real-time echo signal and the cutter head position data are reconstructed by grid to obtain the original polar coordinate radar matrix.

9. A ground-penetrating radar detection system for slurry balance shield tunneling machines, characterized in that, include: The shield cutterhead, recessed mounting base, radar transceiver module, protective packaging components, signal transmission cables, central slip ring assembly, and data processing terminal; The shield cutterhead includes a frame composed of several radially arranged support spokes and a closed panel that is staggered and welded or fixed with bolts between adjacent support spokes; a recessed mounting base is integrally formed or welded to a preset position on the closed panel; a radar transceiver module is embedded in the internal space of the recessed mounting base, and the detection beam emitting surface of the radar transceiver module faces the excavation face; a protective encapsulation component covers the side of the radar transceiver module facing the excavation face. One end of the signal transmission cable is electrically connected to the data interface of the radar transceiver module, and the signal transmission cable extends to the center of the shield cutterhead and connects to the rotating end of the central slip ring assembly; the stationary end of the central slip ring assembly is electrically connected to the data processing terminal. The data processing terminal is used to execute the ground-penetrating radar detection method for slurry balance shield tunnels as described in any one of claims 1 to 8.

10. The ground-penetrating radar detection system for slurry balance shield tunneling according to claim 9, characterized in that, The protective encapsulation assembly includes a wave-transparent waterproof cover and a ceramic wear-resistant plate. The wave-transparent waterproof cover surrounds and seals the front face and side walls of the radar transceiver module, and the ceramic wear-resistant plate is fixed to the outermost surface of the wave-transparent waterproof cover.

Images