Tbm tunneling noise data imaging method based on multiple deep neural networks

Abstract

The application discloses a TBM tunneling noise data imaging method based on a multi-depth neural network, which comprises the following steps: S1, constructing a tunneling noise real-time imaging database; S2, constructing an interference suppression network component based on deep learning; S3, constructing a first arrival wave position prediction network component based on deep learning; S4, constructing an intelligent inversion network component based on deep learning; and S5, TBM tunneling noise data acquisition and intelligent imaging. The application provides high-quality data information for each network component in the subsequent sequence by designing a network structure, thereby assisting in improving the processing precision of the subsequent components, realizing automatic data noise suppression, surrounding rock wave velocity calculation and offset wave velocity field construction, and combining cross-correlation interference and reverse-time migration imaging algorithms to realize rapid processing from tunneling noise observation data to reverse-time migration imaging results, and forming a set of tunnel TBM tunneling noise data real-time intelligent imaging method without manual processing links, which takes into account imaging efficiency and imaging precision.

Description

Technical Field

[0001] This invention belongs to the field of tunnel advanced geological prediction technology, specifically relating to a TBM tunneling noise data imaging method based on multi-depth neural networks. Background Technology

[0002] Tunnel construction faces challenges such as faults, fracture zones, and karst caves, which can easily lead to accidents like machine jamming, water and mud inrushes, and collapses, threatening the safe construction of tunnel projects. Tunnel seismic wave-based early warning methods have the advantages of long detection distances and sensitive responses to structural surfaces in the rock mass, making them an important means of determining the distribution of soil and rock media and the state of adverse geological conditions ahead of the tunnel face.

[0003] TBM cutterhead vibration noise source detection is a method for predicting tunnel seismic waves from noise sources. This method utilizes the noise generated by the TBM cutterhead vibration during excavation as the seismic source. Detectors at the cutterhead record the source waveform data, while detectors deployed on the tunnel sidewalls receive the vibration signals to obtain excavation noise observation data. Through seismic wavefield feature recovery methods such as cross-correlation interferometry, the excavation noise observation data is reconstructed into seismic wave data approximating active source data. Finally, this data is used for forward velocity interface migration imaging to locate faults, fracture zones, and karst caves. Because this method uses TBM excavation vibration noise as the seismic source, it avoids the problem of conventional active source tunnel seismic wave prediction methods requiring detection during TBM shutdown, greatly improving detection efficiency. However, due to limitations in existing seismic data processing algorithms that rely on human experience and computational efficiency, this method currently struggles to achieve high-precision real-time imaging, mainly due to the following three issues:

[0004] Firstly, in the process of restoring the seismic wave field characteristics of tunneling vibration and noise data, the removal of noise interference depends on the experience of the processing personnel. Manual processing is inefficient and its quality is affected by human factors.

[0005] Secondly, when performing migration imaging of data after seismic wavefield feature recovery, the wave velocity of the first arrival wave, which is manually picked up, is often used as the migration wave velocity, i.e., the background wave velocity field during the entire seismic wavefield propagation process. On the one hand, the manual picking process is prone to errors and is inefficient; on the other hand, this method actually uses the surrounding rock wave velocity at the location of the geophones near the tunnel sidewall as the background wave velocity for the entire wavefield propagation, which is an approximation and does not conform to the actual situation, affecting the imaging positioning accuracy.

[0006] Third, if conventional methods such as full waveform inversion of seismic wave velocity are used to obtain the migration velocity model, the inversion iteration process will be time-consuming. Furthermore, due to the presence of coherent noise interference such as virtual phase axes in the cross-correlation data, the strong nonlinearity of the inversion problem, and the inherent problem of the lack of low wavenumber components in small offset data under tunnel observation space, conventional linear inversion methods are prone to getting trapped in local extrema, thus producing erroneous inversion results. Summary of the Invention

[0007] To address the issues of imaging accuracy and efficiency in the aforementioned tunnel seismic wave detection methods based on TBM tunneling noise, this invention proposes a TBM tunneling noise data imaging method based on multiple deep neural networks. This invention intelligently integrates three key aspects of the imaging process that cause the aforementioned problems, and designs three supporting deep neural network components: an interference suppression network component, a first arrival wave position prediction network component, and an intelligent inversion network component. Throughout the process, the network structure is designed so that preceding components provide high-quality data information to subsequent network components, assisting in improving the processing accuracy of subsequent components. This achieves automated data noise suppression, surrounding rock wave velocity calculation, and migration wave velocity field construction. Furthermore, by combining cross-correlation interferometry and inverse time migration imaging algorithms, it enables rapid processing from tunneling noise data to inverse time migration imaging results, forming a real-time intelligent imaging method for tunnel TBM tunneling noise data that balances imaging efficiency and accuracy.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A method for imaging TBM tunneling noise data based on multiple deep neural networks includes the following steps:

[0010] S1. Construct tunnel wave velocity models in batches. For each tunnel wave velocity model, generate corresponding tunneling noise observation data, detector signals at the TBM cutterhead, virtual source shot gather data, first arrival wave labels, uniform wave velocity models, and tunneling noise observation data containing other construction noise through numerical simulation, forming a real-time tunneling noise imaging database.

[0011] S2. Construct a deep learning-based interference suppression network component, and perform self-supervised training on the tunneling noise observation data containing other construction noise as described in step S1, so that it can predict the tunneling noise observation data after noise suppression.

[0012] S3. Construct a deep learning-based first arrival wave position prediction network component, and perform supervised learning training on the network component based on the virtual source shot set data and first arrival wave labels described in step S1, so that it can predict the first arrival wave labels.

[0013] S4. Construct a deep learning-based intelligent inversion network component, and splice the detector signal at the TBM cutterhead described in step S1 with each channel of the tunneling noise observation data described in step S1 along the time axis to obtain spliced ​​tunneling noise data. Based on the spliced ​​tunneling noise data, the tunnel wave velocity model described in step S1, and the uniform wave velocity model described in step S1, supervise the learning training of the network component so that it can predict the tunnel wave velocity model.

[0014] S5. Conduct seismic wave detection and data acquisition of TBM tunneling noise sources in actual tunnel construction projects to obtain measured tunneling noise observation data and measured geophone signals at the TBM cutterhead. Suppress the measured tunneling noise observation data through the interference suppression network component described in step S2 to predict the measured tunneling noise observation data after noise suppression. Obtain the measured virtual source shot gather data by cross-correlation interference between this data and the measured geophone signals at the TBM cutterhead.

[0015] The measured virtual source shot collection data is input into the first arrival wave position prediction network component to predict the measured first arrival wave label, and the first arrival wave velocity is calculated based on the first arrival wave slope in the measured first arrival wave label.

[0016] Based on the initial arrival wave velocity, a uniform wave velocity model is constructed. The uniform wave velocity model, the measured detector signal at the TBM cutterhead, and the spliced ​​measured tunneling noise observation data obtained by splicing each channel along the time axis are used as inputs to the intelligent inversion network component to predict the tunnel wave velocity model.

[0017] Reverse-time migration imaging was performed on the wave velocity model of the probe tunnel using measured virtual source shot gather data.

[0018] In step S1, the tunnel seismic wave detection and observation method using passive source wave field forward modeling and conventional TBM tunneling noise as the source is adopted to generate tunneling noise observation data corresponding to each velocity model, and the virtual source shot gather data corresponding to each tunneling noise observation data is generated by cross-correlation interferometry.

[0019] Active source forward modeling is performed using the same observation method as passive source wavefield forward modeling, and the first arrival wave initiation point position is marked based on the active source forward modeling data to generate first arrival wave labels;

[0020] A uniform wave velocity model with the same size is constructed using the wave velocity at the tunnel surrounding rock in the tunnel wave velocity model, i.e., the first arrival wave velocity. In addition, the actual construction noise measured in the tunnel is added to each tunneling noise observation data. The above multiple sets of corresponding data and wave velocity models form a real-time imaging database of tunneling noise.

[0021] In step S2, the interference suppression network component adopts an autoencoder structure, wherein the encoder part is composed of multiple hollow convolutional layers. During the network training process, the input is the tunneling noise observation data containing other construction noises from the real-time tunneling noise imaging database, and the input data is used as the label data for network training.

[0022] In step S3, the input to the first arrival wave position prediction network component during training is the virtual source shot set data corresponding to each tunneling noise observation data in the real-time tunneling noise imaging database obtained in step S1.

[0023] The label data for network training consists of first-arrival labels from the real-time imaging database of tunneling noise, corresponding to the first-arrival start point positions of the active source forward modeling data.

[0024] In step S4, the intelligent inversion network component includes a tunneling noise observation data encoder and a wave velocity model construction network module. The tunneling noise observation data encoder consists of multiple sets of hollow convolutional layers. The input is the spliced ​​tunneling noise data, and the output is a feature vector v. d ;

[0025] The input to the wave speed model construction network module is a uniform wave speed model. Each network layer in this module is followed by a cross-attention mechanism layer, and the input to each cross-attention mechanism layer is the feature vector v. d Compared to the output of the intermediate layers preceding this layer, the network structure of the intelligent inversion network component is represented as follows:

[0026]

[0027] Where, d c This represents the spliced ​​tunneling noise data; m0 represents the uniform wave velocity model; w3 represents the network parameters of the intelligent inversion network component; En d This represents the encoder for tunneling noise observation data; VBnet represents the wave velocity model construction network module; m represents the predicted wave velocity model, i.e., the offset wave velocity model.

[0028] The cross-attention mechanism layer is represented as follows:

[0029]

[0030] in,

[0031]

[0032] In the formula, s v Represents the eigenvector v d Length, W Q With W KBoth are network parameter matrices, where F(m0,w3) represents the output of the intermediate layer of the network before the cross-attention mechanism layer in VBnet;

[0033] The uniform wave velocity model from the real-time tunneling noise imaging database and the stitched tunneling noise data are used as inputs, and the tunnel wave velocity model is used as a label to supervise the training of the intelligent inversion network component.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] First, this invention addresses the problem of data processing efficiency in current TBM tunneling noise source seismic wave tunnel detection algorithms by using deep neural networks to intelligently process multiple steps in the process. After all components are built, the entire imaging process eliminates time-consuming computational processes such as inversion iterations, and directly implements processes that require manual handling in conventional detection through algorithms. This significantly improves algorithm efficiency compared to traditional processing methods, and promises to achieve real-time imaging of adverse geological features after acquiring detection data.

[0036] Second, this invention addresses the problem affecting imaging accuracy in current TBM tunneling noise source seismic wave tunnel detection algorithms, namely, the simplification of the wave velocity calculation process. This invention inverts wave velocity based on both first-arrival wave velocity and detection data. It utilizes a uniform wave velocity model to provide low wavenumber information and alleviate inversion nonlinearity issues, while leveraging the rich full-waveform information in the tunneling noise data to provide high-frequency information for accurate reconstruction of the wave velocity distribution ahead of the tunnel face. This process incorporates a cross-attention mechanism layer, a structure commonly used to enhance the semantic correlation between two input sequences, which is particularly useful for wave velocity inversion tasks. In this context, the background wave velocity field information from the uniform wave velocity model and the high-frequency amplitude variation information contained in the tunneling noise data correspond to the high-wavenumber wave velocity interface information and the low-wavenumber background wave velocity field distribution information contained in the wave velocity inversion results, respectively. By mining the complementary wave velocity information in the two datasets and matching them with the cross-attention mechanism, a mapping from tunneling noise data to wave velocity distribution is directly established through a deep learning network, enabling the rapid acquisition of more accurate wave velocity information. Combined with the virtual source shot gather data obtained after suppressing interference in the tunneling noise observation data, more accurate imaging and positioning of adverse geological bodies can be achieved. Attached Figure Description

[0037] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0038] Figure 1 This is a flowchart of the TBM tunneling noise data imaging method based on multiple deep neural networks of the present invention;

[0039] Figure 2This is an architecture diagram of the TBM tunneling noise data imaging method based on multiple deep neural networks of the present invention;

[0040] Figure 3 This is a schematic diagram of the velocity model and the tunneling noise observation data, the noisy virtual source shot gather data, the virtual source shot gather data and the first arrival wave label generated by numerical simulation used in this invention;

[0041] Figure 4 This is a schematic diagram of the predicted tunnel wave velocity model and the RTM imaging results of the virtual source shot gather according to the present invention. Detailed Implementation

[0042] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0044] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0045] Specifically, the method flowchart provided in this embodiment is as follows: Figure 1 As shown, the overall architecture of the method based on three network components is as follows: Figure 2 As shown, it includes the following steps:

[0046] Step S1: Construct 10,000 two-dimensional tunnel wave velocity models based on the geological conditions of the tunnel to be detected. Each wave velocity model includes wave velocity interfaces where abrupt changes occur, representing adverse geological bodies such as faults and fracture zones. The model mesh size is dx = dz = 1m, and the model dimensions are z × x = 80 × 150 grids. A 20-grid thick sponge-absorbing boundary is also set around the model. The tunnel is positioned at the exact center of the z-axis, with a width of 10m and a length of 50m in the x-direction. The wave velocity is uniformly set to 340m / s. The surrounding rock wave velocity ranges from 1000m / s to 5000m / s. A uniform wave velocity model of the same size is generated for each model based on the tunnel surrounding rock wave velocity. A survey line is laid on both the upper and lower sides of the tunnel wall. Each survey line contains 3 seismic source points and 6 geophone points. The seismic source points are evenly distributed at a distance of 2m to 10m from the working face, with a spacing of 4m. Geophones were positioned 11m to 31m from the working face, with a spacing of 4m. The source-loaded waveform of the tunneling noise observation data was obtained by convolution of the Ricker wavelet with a random time series. A 300Hz Ricker wavelet source was used for excitation. The time step recorded by the detector was 1ms, and the total duration was 0.4s. Forward modeling was performed using the second-order spatial and second-order temporal finite difference method of the constant density acoustic wave equation to obtain tunneling noise observation data with a size of s×r×t=3×6×400. The detector signal at the TBM cutterhead was also obtained with a size of s×r×t=3×1×400. The noise data was restored to a virtual source shot gather with a size of s×r×t=3×6×200 using the cross-correlation interferometry method. A three-dimensional matrix with the same size as the virtual source shot gather data was constructed. Based on the first arrival wave start point position of the active source data with the same observation method as the virtual source shot gather, positions with a time smaller than the first arrival wave start point were marked as 0 in the label matrix, and those larger were marked as 1. Values ​​were assigned to the three-dimensional matrix to obtain the first arrival wave labels. Gaussian noise was added to each tunneling noise observation data to ensure the signal-to-noise ratio was between -10 dB and 15 dB. The two-dimensional constant-density acoustic wave equation can be expressed as:

[0047]

[0048] In the formula, P represents the pressure wave field, m represents the wave velocity model, and s represents the source term. A total of 10,000 sets of two-dimensional tunnel wave velocity models and corresponding data types were obtained, which were divided into a training set of 8,000 sets, a validation set of 1,000 sets, and a test set of 1,000 sets, in a ratio of 8:1:1. The training and validation sets were used for training the network parameters of the three components in subsequent steps S2, S3, and S4, while the test set was used to test the effectiveness of the entire method in step S5.

[0049] Of course, in other embodiments, multiple wave velocity models can be established using other data. Alternatively, during the model establishment process, the selected parameters may not be the same as those provided in the above embodiments and can be transformed.

[0050] In other embodiments, wavefield simulation is performed for each wave velocity model with fixed source, detector location and observation time, and wavefield data is recorded at the detector location to obtain seismic observation data corresponding to the geological wave velocity model.

[0051] In this embodiment, the database contains a tunnel wave velocity model, a uniform wave velocity model constructed from surrounding rock wave velocities, corresponding tunneling noise observation data, virtual source shot gather data, tunneling noise observation data with added noise, and first arrival wave labels, as follows: Figure 3 As shown.

[0052] Step S2: Construct the interference suppression network component. The input during network training is noisy tunneling noise observation data from the real-time tunneling noise imaging database, and this input data is used as the label data for network training. The network consists of 6 cascaded dilated convolutional layers and 6 cascaded deconvolutional layers. The interference suppression network component is trained using data from the dataset, and the objective function can be expressed as:

[0053]

[0054] in, The data represents the tunneling noise observation data, which also serves as the label for network training; w1 represents the network parameters of the convolutional autoencoder network; and Denoi represents the interference suppression network components. A total of 100 training epochs were conducted using the Adam optimizer with a learning rate of 1×10⁻⁶. -4 The batch size is 8. This process is based on the disordered and random distribution of interference noise in the data and the image regularization ability of dilated convolution itself. Noise suppression is achieved through unsupervised training. Dilated convolution has a larger receptive field than convolutional networks, making it more suitable for long-term sampling of tunneling noise observation data.

[0055] Step S3: Construct the first-arrival position prediction network component. The input to this network during training is the virtual source shot gather data corresponding to each tunneling noise observation data in the real-time tunneling noise imaging database, with a size of s×r×t=3×6×200. The label data is the first-arrival labels from the real-time tunneling noise imaging database, also with a size of 3×6×200. The first-arrival position prediction network component consists of 6 cascaded hollow convolutional layers, 6 cascaded deconvolutional layers, and skip connections between each convolutional layer and each deconvolutional layer. The objective function for the training process can be expressed as:

[0056]

[0057] Where, d v w2 represents the input virtual source shot set data; w2 represents the network parameters of the first arrival position prediction network component; FPnet represents the first arrival position prediction network component. This represents the initial arrival wave label. A total of 100 training rounds were conducted using the Adam optimizer with a learning rate of 5×10⁻⁶. -5 The batch size is 32.

[0058] Step S4: Construct an intelligent inversion network component. This component stitches the detector signal at the TBM cutterhead in the real-time tunneling noise imaging database with each channel of the tunneling noise observation data along the time axis to obtain the stitched tunneling noise data. This network component has two inputs: the stitched tunneling noise data and its corresponding uniform wave velocity model. The output is a predicted wave velocity model, which is the migration wave velocity model used in subsequent migration imaging. The training labels use the tunnel wave velocity model from the real-time tunneling noise imaging database. This component includes a tunneling noise observation data encoder and a wave velocity model construction network module. The tunneling noise observation data encoder consists of multiple sets of hollow convolutional layers. The tunneling noise observation data encoder compresses the tunneling noise observation data into a feature vector v. d In the wave speed model construction network module, each network layer is followed by a cross-attention mechanism layer, and the input of each cross-attention mechanism layer is v. d Compared to the output of the intermediate layers preceding this layer, the network structure of the intelligent inversion network component is represented as follows:

[0059]

[0060] Where, d c This represents the spliced ​​tunneling noise data; m0 represents the uniform wave velocity model; w3 represents the network parameters of the intelligent inversion network component; En d This represents the encoder for tunneling noise observation data; VBnet represents the wave velocity model construction network module; m represents the predicted wave velocity model, i.e., the offset wave velocity model.

[0061] The objective function for network training can be expressed as:

[0062]

[0063] In the formula, This represents the labeled data. A total of 200 training rounds were conducted, using the Adam optimizer with a learning rate of 1×10⁻⁶. -4 The batch size is 16.

[0064] Step S5: In practical application, the same observation method as in this embodiment for generating simulated data is used to collect TBM tunneling noise data from actual tunnel construction projects. This embodiment uses data generated through numerical simulation from the test set to replace the actual TBM tunneling noise data from tunnel construction projects for testing the entire method. A set of tunneling noise observation data from the test set ( Figure 3(As shown) Input the trained interference suppression network component, output the interference-suppressed data, and perform cross-correlation interference with the detector data at the cutterhead source to obtain virtual source shot gather data. Input the first arrival wave position prediction network component to predict the first arrival wave label, and calculate the first arrival wave velocity according to the slope of the first arrival wave marked in the data using the following formula:

[0065] v = x r / x t

[0066] Where x r To predict the actual distance represented by the detector's r-direction coordinate axis in the data, x t To predict the actual duration represented by the time scale t-axis in the data. Figure 3 The data x shown r For 20m, x t If the first arrival time is 20ms, then the initial arrival wave velocity v is 1000m / s. After obtaining the initial arrival wave velocity, a uniform wave velocity model with the same dimensions as the wave velocity model used in training the network is constructed. The stitched tunneling source data and the corresponding uniform wave velocity model are input into the intelligent inversion network component to obtain the predicted wave velocity model, i.e., the migration wave velocity model. Finally, using a virtual source shot gather and the migration wave velocity model, reverse time migration imaging (RTM) is performed based on normalized cross-correlation imaging conditions to achieve rapid imaging of the wave velocity interface in front of the tunnel.

[0067]

[0068] Where S is the forward propagation wavefield, R is the reverse propagation residual wavefield, I is the imaging result, x represents the coordinates, and t represents the time step of the seismic data.

[0069] The total processing time for this set of training data was less than 2 minutes. Figure 4 The reverse time migration imaging results shown in the figure reflect the stratification of wave velocity in front of the tunnel face well, indicating that the method of the present invention can accurately and quickly locate the location of adverse geological bodies in front of the tunnel face.

[0070] The program used in this embodiment is implemented based on the PyTorch library in the Python language and runs on four NVIDIA TITAN RTX graphics cards with 48GB of video memory.

[0071] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0072] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0073] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for imaging TBM tunneling noise data based on multiple deep neural networks, characterized in that, Includes the following steps: S1. Construct tunnel wave velocity models in batches. For each tunnel wave velocity model, generate corresponding tunneling noise observation data, detector signals at the TBM cutterhead, virtual source shot gather data, first arrival wave labels, uniform wave velocity models, and tunneling noise observation data containing other construction noise through numerical simulation, forming a real-time tunneling noise imaging database. S2. Construct a deep learning-based interference suppression network component, and perform self-supervised training on the tunneling noise observation data containing other construction noise as described in step S1, so that it can predict the tunneling noise observation data after noise suppression. S3. Construct a deep learning-based first arrival wave position prediction network component, and perform supervised learning training on the network component based on the virtual source shot set data and first arrival wave labels described in step S1, so that it can predict the first arrival wave labels. S4. Construct a deep learning-based intelligent inversion network component, and splice the detector signal at the TBM cutterhead described in step S1 with each channel of the tunneling noise observation data described in step S1 along the time axis to obtain spliced ​​tunneling noise data. Based on the spliced ​​tunneling noise data, the tunnel wave velocity model described in step S1, and the uniform wave velocity model described in step S1, supervise the learning training of the network component so that it can predict the tunnel wave velocity model. S5. Conduct seismic wave detection and data acquisition of TBM tunneling noise sources in actual tunnel construction projects to obtain measured tunneling noise observation data and measured geophone signals at the TBM cutterhead. Suppress the measured tunneling noise observation data through the interference suppression network component described in step S2 to predict the measured tunneling noise observation data after noise suppression. Obtain the measured virtual source shot gather data by cross-correlation interference between this data and the measured geophone signals at the TBM cutterhead. The measured virtual source shot collection data is input into the first arrival wave position prediction network component to predict the measured first arrival wave label, and the first arrival wave velocity is calculated based on the first arrival wave slope in the measured first arrival wave label. Based on the initial arrival wave velocity, a uniform wave velocity model is constructed. The uniform wave velocity model, the measured detector signal at the TBM cutterhead, and the spliced ​​measured tunneling noise observation data obtained by splicing each channel along the time axis are used as inputs to the intelligent inversion network component to predict the tunnel wave velocity model. Reverse-time migration imaging was performed on the wave velocity model of the probe tunnel using measured virtual source shot gather data.

2. The TBM tunneling noise data imaging method based on multiple deep neural networks according to claim 1, characterized in that, In step S1, the tunnel seismic wave detection and observation method using passive source wave field forward modeling and conventional TBM tunneling noise as the source is adopted to generate tunneling noise observation data corresponding to each velocity model, and the virtual source shot gather data corresponding to each tunneling noise observation data is generated by cross-correlation interferometry. Active source forward modeling is performed using the same observation method as passive source wavefield forward modeling, and the first arrival wave initiation point position is marked based on the active source forward modeling data to generate first arrival wave labels; A uniform wave velocity model with the same size is constructed using the wave velocity at the tunnel surrounding rock in the tunnel wave velocity model, i.e., the first arrival wave velocity. In addition, the actual construction noise measured in the tunnel is added to each tunneling noise observation data. The above multiple sets of corresponding data and wave velocity models form a real-time imaging database of tunneling noise.

3. The TBM tunneling noise data imaging method based on multiple deep neural networks according to claim 1, characterized in that, In step S2, the interference suppression network component adopts an autoencoder structure, wherein the encoder part is composed of multiple hollow convolutional layers. During the network training process, the input is the tunneling noise observation data containing other construction noises from the real-time tunneling noise imaging database, and the input data is used as the label data for network training.

4. The TBM tunneling noise data imaging method based on multiple deep neural networks according to claim 1, characterized in that, In step S3, the input to the first arrival wave position prediction network component during training is the virtual source shot set data corresponding to each tunneling noise observation data in the real-time tunneling noise imaging database obtained in step S1. The label data for network training consists of first-arrival labels from the real-time imaging database of tunneling noise, corresponding to the first-arrival start point positions of the active source forward modeling data.

5. The TBM tunneling noise data imaging method based on multiple deep neural networks according to claim 1, characterized in that, In step S4, the intelligent inversion network component includes a tunneling noise observation data encoder and a wave velocity model construction network module. The tunneling noise observation data encoder consists of multiple sets of hollow convolutional layers. The input is the spliced ​​tunneling noise data, and the output is a feature vector v. d ; The input to the wave speed model construction network module is a uniform wave speed model. Each network layer in this module is followed by a cross-attention mechanism layer, and the input to each cross-attention mechanism layer is the feature vector v. d Compared to the output of the intermediate layers preceding this layer, the network structure of the intelligent inversion network component is represented as follows: Where, d c This represents the spliced ​​tunneling noise data; m0 represents the uniform wave velocity model; w3 represents the network parameters of the intelligent inversion network component; En d This represents the encoder for tunneling noise observation data; VBnet represents the wave velocity model construction network module; m represents the predicted wave velocity model, i.e., the offset wave velocity model. The cross-attention mechanism layer is represented as follows: in, In the formula, s v Represents the eigenvector v d Length, W Q With W K Both are network parameter matrices, where F(m0,w3) represents the output of the intermediate layer of the network before the cross-attention mechanism layer in VBnet; The uniform wave velocity model from the real-time tunneling noise imaging database and the stitched tunneling noise data are used as inputs, and the tunnel wave velocity model is used as a label to supervise the training of the intelligent inversion network component.

Images