Tunnel seismic data reconstruction method based on mixed attention unet neural network

Abstract

The application provides a tunnel seismic data reconstruction method based on a mixed attention Unet neural network, which automatically learns the characteristics between existing tunnel seismic data through an artificial intelligence neural network, and establishes a nonlinear mapping relationship between missing data and complete data; through the Unet neural network in the artificial intelligence, which has many powerful functions such as image semantic segmentation and communication signal denoising, a mixed attention mechanism based on channel attention and spatial attention is designed and integrated into the skip link part of the Unet network, so that the network model can extract effective signal events and discard redundant information such as noise; and through classification of the measured data, a strong generalization training data set and a test data set are established, the network model parameters are trained and the effect is verified, the training parameters of the model with the highest accuracy of the tunnel seismic data reconstruction result are taken as the final result, so that the intelligent and accurate reconstruction of missing data in the tunnel seismic wave method advanced geological prediction is realized.

Description

Technical Field

[0001] This invention belongs to the fields of engineering geophysics and artificial intelligence technology, and specifically relates to a method for reconstructing advanced geological prediction data using the Unet neural network tunnel seismic wave method incorporating a hybrid attention mechanism. Background Technology

[0002] Deeply buried, long hydraulic tunnels face complex topographic, hydrogeological, and engineering geological conditions, poor surrounding rock stability, high ground stress environments, and significant variations in rock mass types. These challenges include crossing active faults, water and mud inrushes due to high external water pressure, water environment impacts, and rock bursts in hard rock caused by high ground stress in deep-buried tunnels. During the construction of hydraulic tunnels and underground engineering projects, the degree of understanding and mastery of the geological conditions ahead of the tunnel face is one of the key factors ensuring rapid and safe construction (Hao Junsuo, 2021). Due to the complex and variable geological composition of the surrounding rock in hydraulic tunnels and the limitations of exploration methods, it is difficult to accurately determine the location, nature, and scale of adverse geological bodies before construction. Advanced geological prediction is a key technology for solving these problems and has received attention and widespread acclaim from the tunnel engineering community worldwide. It is hoped that the prediction results can provide information on the mechanical properties and structural characteristics of the rock mass, enabling a better evaluation of the stability of the tunnel's surrounding rock. The tunnel seismic wave method is currently the most important prediction method. Compared with geophysical prediction methods such as induced polarization method, resistivity method, and transient electromagnetic method, it can calculate the shear wave velocity of rock mass, providing an effective way to obtain information on the dynamic parameters of surrounding rock in hydraulic tunnels.

[0003] In terms of the layout of the tunnel seismic wave method for advanced geological prediction acquisition and observation systems, there are currently three main approaches: point layout, linear layout, and spatial layout. The land sonar method (Zhong Shihang, 1995) uses a point-based observation layout, employing a self-excited and self-received observation system on the tunnel face for advanced tunnel detection. Zeng Zhaohuang et al. (1994) proposed the negative velocity method (TVSP), which involves arranging excitation and receiving points on the tunnel sidewalls and inferring the location of anomalous structures ahead of the tunnel from the intersection of the reflected wave phase axis curve and the direct wave phase axis curve. The TSP (Tunnel Seismic Prediction) tunnel seismic prediction system proposed by Dickmann (1996) uses a similar observation method to the negative velocity method, the main difference being the use of diffraction scanning and superimposed migration imaging for data processing. Subsequently, Liu Yunzhen (2009) developed the TGP12 multifunctional tunnel geological advanced prediction instrument, which, compared to TSP, adds velocity scanning and depth migration imaging. Zhou Liming (2011) numerically analyzed the mechanism characteristics of elastic waves propagating in tunnel surrounding rock, including reflection, scattering, and propagation attenuation. He also simulated the elastic wave field of an adverse geological body theoretical model using finite difference forward modeling in a tunnel environment containing air, providing a theoretical basis for the design of observation systems, data processing, and data interpretation for this type of method. Li Shucai (2014) argued that linear observation system layouts, represented by TSP and negative apparent velocity methods, suffer from inaccurate wave velocity estimation and unreliable detection results. Zhou Liming (2016) compared TSP data with the geological results revealed by the 3D prediction results of the Wudongde flood discharge tunnel. Fu Daiguang (2016) analyzed the influence of boreholes, receiver holes, and borehole-receiver hole relationships on prediction results in TSP data acquisition through prediction examples of multiple tunnels on the Anjiang Expressway, and listed successful prediction cases of karst caves, loose fracture zones, and soft interlayers using TSP. Zhao Yonggui (2006) proposed TST (Tunnel Seismic Tomography) tunnel seismic CT imaging technology, which also adopts a spatial array observation system layout and uses visualization imaging technology to vividly show the detection situation within a range of 150m in front of the tunnel. Internationally, Inazaki et al. (1999) used the horizontal acoustic profiling method (HSP), arranging the excitation and receiver points on both sides of the tunnel. Neil et al. (1999) used Tunnel Reflection Tomography (TRT) to achieve three-dimensional spatial observation of the tunnel, fully obtaining spatial wavefield information. However, this type of detection method requires the tunnel face and both side walls, is time-consuming, and uses a hammer-driven seismic source. In cases of poor surrounding rock conditions, the signal attenuates rapidly, making long-distance prediction difficult. In recent years, with the widespread application of TBM tunneling technology, advanced geological prediction technology under specific TBM environments faces greater challenges. Traditional seismic and electromagnetic prediction methods are challenged due to environmental interference and spatial limitations.The Integrated Seismic Imaging System (ISIS) developed by GFZ in Germany (Borm, 2003) uses a TBM as the source to generate seismic waves and Fresnel migration imaging technology for data processing. Petronio and Poletto (2008) proposed a drilling-while-drilling seismic prediction method applicable to TBM-constructed tunnels, utilizing multi-source seismic waves generated by the TBM cutterhead cutting rock as excitation signals to achieve advanced geological prediction without shutting down the machine. In 2008, Li Cangsong et al. first proposed the passive source HSP (horizontal sonic seismic profiling) method, which uses the vibration signal generated by the TBM cutterhead shearing rock as the excitation source. Lu Song et al. (2019) optimized the hardware and software design of the HSP system to make it miniaturized, automated, and intelligent. Liu Jingang et al. (2021) used HSP technology for advanced geological prediction and mounted it on a TBM to obtain stratigraphic characteristic parameters in front of the tunnel face.

[0004] In the processing of tunnel seismic wave method data for advanced geological prediction, the main techniques include denoising, wavefield separation, and migration imaging. The denoising process is similar to conventional seismic data processing, with commonly used methods including Fourier transform and wavelet transform. Methods for separating P-waves and S-waves include Radon transform based on kinematic characteristics, median filtering, and FK domain filtering, polarization filtering, singular value decomposition, divergence, and curl filtering based on dynamic characteristics. Migration imaging mainly includes ray migration based on Huygens' principle and wave equation migration imaging based on wave theory. Aashida (2002) borrowed the Kirchhoff pre-stack time migration principle from reflection seismic data and used the isochronous surface method for time-domain imaging. Ray-based diffraction stacking migration imaging, which uses the isochronous surface theory, has difficulty resolving complex geological structures. Luth et al. (2008) first applied Kirchhoff pre-stack depth migration to tunnel advanced detection, successfully predicting adverse geological bodies in complex rock environments. Tzavaras et al. (2012) proposed Fresnel volume migration imaging and reflection spectrum imaging for data processing to improve resolution. Liu Shengdong et al. (2015) introduced polarization coefficient, principal polarization direction, and coherence factor to propose a spatially independent dynamic polarization migration imaging method for reflected waves to eliminate symmetry anomaly artifacts and improve spatial resolution. Jin Hongdi (2012) verified the feasibility of applying common scattering point gather migration technology to tunnel advance detection through numerical simulation. Jannis (2012) used Kirchhoff pre-stack depth migration method to process tunnel seismic data. Bai Mingzhou et al. (2016) optimized the three-dimensional geological structure imaging technology based on the TRT system by improving the calculation method of direct wave velocity and background wave velocity. Chang Xu et al. (2006) used the zero-time imaging condition of the wave equation to image zero-offset seismic records; Liu Jiangping et al. (2014) realized two-dimensional reverse time migration imaging of tunnels based on the elastic wave equation based on the excitation time imaging condition. Wang Jing (2017) conducted research on a three-dimensional P-wave reverse-time migration imaging method for tunnels based on elastic wave reverse-time extrapolation. The results showed that the method has certain noise resistance and can improve imaging quality. Cheng Luming (2016) conducted research on a multi-parameter three-dimensional imaging method for tunnel elastic waves using the principles of seismic wave diffraction scanning migration superposition and common reflection surface element superposition. Wang Xu et al. (2019) studied the applicability of multi-source seismic interferometry in tunnel advanced geological prediction and discussed the main influencing factors of its imaging results.

[0005] However, in actual tunnel geological prediction using seismic wave methods, some effective seismic traces are not received due to detector malfunctions and data loss in the acquisition software. Simultaneously, strong noise interference in some tunnel environments or poor coupling quality between the detector and the rock mass results in low signal-to-noise ratios in the received data, interfering with velocity inversion and structural plane imaging, necessitating the manual removal of this data. These two reasons reduce the amount of effective data for velocity inversion and structural plane imaging of the rock mass ahead of the tunnel face. Currently, traditional data reconstruction methods based on fx deconvolution and Randon transform rely on theoretical model assumptions, and their reconstruction effectiveness is significantly affected by noise and the consistency of effective signals. Therefore, there is an urgent need to develop a method based on nonlinear mapping using artificial neural networks to effectively reconstruct missing tunnel seismic data and increase the amount of effective data for tunnel seismic wave geological prediction. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a Unet neural network-based method for reconstructing advanced geological prediction data using tunnel seismic wave methods, which incorporates a hybrid attention mechanism. This method automatically learns the characteristics between existing tunnel seismic data through an artificial intelligence neural network and establishes a nonlinear mapping relationship between missing and complete data, thereby solving the problem of accurately reconstructing advanced geological prediction data using tunnel seismic wave methods.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for reconstructing tunnel seismic data incorporating a hybrid attention Unet neural network includes the following steps:

[0009] a. Design a channel attention mechanism module, which consists of three parts: squeezing, activation, and channel attention calculation. The squeezing part compresses the feature information output by the convolutional layer by performing global average pooling or max pooling on the features by channel. The activation part includes Conv1-Act1-Conv2-Act2, where Conv represents the convolutional layer and Act represents the activation layer. The channel attention calculation part multiplies the weight values ​​with the feature maps output by the convolutional layer to achieve weighted processing of the channels, thereby achieving attention to important features.

[0010] b. Design a spatial attention mechanism module, which scores feature data in the spatial domain, and channel attention, which scores feature data in the channel domain.

[0011] c. Design a hybrid attention mechanism module, which is composed of the channel attention mechanism module and the spatial attention mechanism module mentioned above, combined by a Concat operation. The features from the convolutional combination layer are simultaneously input into the channel attention mechanism module and the spatial attention mechanism module. The features output by the channel attention mechanism module and the features output by the spatial attention mechanism module are combined by a Concat operation, and then activated by convolution, batch normalization and ReLU function. The feature is restored to the same size as the input feature map, which is the output feature of this hybrid attention module.

[0012] d. Design a Unet neural network architecture that incorporates a hybrid attention mechanism. The Unet neural network and the hybrid attention module are built using the PyTorch machine learning library. The Unet neural network includes convolutional layers, downsampling layers, upsampling layers, and skip connections. By adding the hybrid attention mechanism module to the skip connection part of the regular Unet neural network, the hybrid attention mechanism can be integrated.

[0013] e. Design a loss function for the Unet neural network that incorporates a hybrid attention mechanism. This loss function J SAUnet It can be represented as:

[0014] J SAUnet =J MSE +λJ Attention

[0015] Among them, J MSE Let J be the mean squared error loss function. Attention λ represents the loss function constraint term of the hybrid attention module, and λ represents the coefficient of the adjustment constraint term;

[0016] f. Classify the actual tunnel seismic wave method advanced geological prediction data, establish training and test datasets, extract complete data with high signal-to-noise ratio, and manually sample data with random trace missing, continuous trace missing, and uniform trace missing to generate training data. The corresponding complete data are used as labels to establish the training dataset Tr. set The missing track numbers in the actual data of missing tracks that have been excavated and verified for rock mass properties are padded with zeros to serve as test data (Te). set The set is used to verify the effectiveness of the data reconstruction results;

[0017] g. Initialize network parameters, determine the learning rate, learning rate decay rate, number of input samples per cycle, and number of learning iterations;

[0018] h, Tr set Using a sliding time window W s T Slicing the data into pieces yields the corresponding training data slices. Where T represents the maximum number of tunnel seismic data, and s represents the step size of each movement of the sliding window, i.e., the sampling interval of the seismic data;

[0019] i. will Input the Unet neural network parameters with a hybrid attention mechanism, adjust the hyperparameters and data volume, and perform multiple training sessions. Compare the output results with the labels using FK spectrum and single-channel waveform analysis to determine the accuracy of the data reconstruction results. Take the network model parameters with the minimum loss function and the highest data reconstruction accuracy as the final training result.

[0020] j. Te set The data reconstruction test was performed on the trained Unet neural network model with a hybrid attention mechanism. The velocity inversion and structural surface imaging were performed on the reconstructed complete data, and the accuracy was compared with the excavation verification results.

[0021] Further, step b, the evaluation process, specifically involves: first, performing global average pooling on the feature data along the channel dimension; then, passing it through a convolutional layer to score the importance of the features in space; next, using the Sigmoid function to limit the scoring range to (0,1) to obtain the weights of each spatial feature; finally, multiplying the above spatial weights with the input features of the module to achieve attention to important spatial features.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] 1. While leveraging the advantages of conventional Unet neural networks in extracting high-dimensional features and performing data recovery, a hybrid attention mechanism is proposed and designed and integrated into the skip link part of the Unet neural network. In the multi-scale feature map fusion part, spatial attention and channel attention are used to improve the attention to the in-phase axis of the limited reflection signal in tunnel seismic data. At the same time, useless information such as noise is discarded to improve the accuracy and signal-to-noise ratio of the data reconstruction results.

[0024] 2. A loss function based on the mean square error of network output and label was adopted, and the mean square error of feature maps before and after hybrid attention output was used as a constraint term to enhance the effect of the hybrid attention mechanism and improve the accuracy of the network output results, i.e., the reconstructed tunnel seismic data.

[0025] 3. High-quality real tunnel seismic data is used to generate a missing tunnel training dataset through manual sampling, which makes the data features learned by the network more closely resemble the actual data and can improve the generalization of network applications. Attached Figure Description

[0026] Figure 1 Channel attention module structure diagram;

[0027] Figure 2 Spatial attention module structure diagram;

[0028] Figure 3 Hybrid attention module structure diagram;

[0029] Figure 4 A diagram of the Unet neural network architecture incorporating a hybrid attention mechanism;

[0030] Figure 5a Missing seismic data for tunnels;

[0031] Figure 5b Tunnel seismic data reconstructed using a Unet neural network incorporating a hybrid attention mechanism;

[0032] Figure 5c FK spectrum of the label data;

[0033] Figure 5d Figure 5b The FK spectrum of the reconstructed data is shown. Detailed Implementation

[0034] The Unet neural network tunnel seismic wave method for advanced geological prediction data reconstruction, which incorporates a hybrid attention mechanism, was implemented on the Anaconda platform using the Python compiler. The program was written in the PyTorch machine learning library framework, and a CUDA parallel computing architecture was built according to the corresponding computational requirements. The graphics card used was an RTX 4090.

[0035] The core of this approach is to propose a Unet neural network-based method for reconstructing tunnel seismic wave-based advanced geological prediction data, incorporating a hybrid attention mechanism. In actual tunnel seismic wave-based advanced geological prediction, some effective seismic channels are not received due to detector malfunctions or data loss in the acquisition software. Simultaneously, strong noise interference in certain tunnel environments or poor coupling quality between the detector and the rock mass results in low signal-to-noise ratios in the received data, interfering with velocity inversion and structural plane imaging, necessitating the manual removal of this data. These data gaps reduce the amount of effective data for velocity inversion and structural plane imaging of the rock mass ahead of the tunnel face. Therefore, this approach utilizes the Unet neural network, which possesses powerful functions in artificial intelligence such as image semantic segmentation, communication signal denoising, and geophysical inversion. A hybrid attention mechanism based on channel attention and spatial attention is designed and integrated into the skip link portion of the Unet network to improve the network model's extraction of effective signal phase axes while discarding redundant information such as noise. By classifying the measured data to establish training and testing datasets, the network model parameters were trained and their effectiveness was verified. The model training parameters with the highest accuracy in the tunnel seismic data reconstruction results were taken as the final result.

[0036] This invention incorporates a hybrid attention Unet neural network method for tunnel seismic data reconstruction. It designs a hybrid attention mechanism based on channel attention and spatial attention, integrating it with the skip link part of the Unet network. This improves the network model's ability to extract effective signal phase axes while discarding redundant information such as noise. Furthermore, it establishes highly generalizable training and testing datasets by classifying measured data, enabling intelligent and accurate reconstruction of missing data from tunnel seismic wave-based advanced geological prediction. The method includes the following steps:

[0037] a. Design of the channel attention mechanism module. The channel attention mechanism module consists of three parts: compression, activation, and channel attention calculation. The compression part compresses the H*W*C dimensional feature information output by the convolutional layer into a 1*1*C dimensional vector by performing global average pooling or max pooling on the features by channel. Here, H and W represent the height and width of the data, respectively, and C is the number of feature channels. The activation part includes Conv1-Act1-Conv2-Act2, where Conv represents the convolutional layer and Act represents the activation layer. Conv1 reduces the feature dimension to 1 / r of the original, where r is equal to the reciprocal of the number of one-dimensional convolutional kernels; Act1 is a ReLU function to achieve non-linearity of the data; Conv2 increases the dimension back to the original size of 1*1*C; Conv2 is a Sigmoid function, the output of which is a 1*1*C weight, with a value between (0,1). The channel attention calculation part multiplies the weight value with the feature map output by the convolutional layer to achieve weighted processing of the channels, thereby achieving attention to important features. Figure 1 );

[0038] b. Design a spatial attention mechanism module. The spatial attention mechanism module scores feature data in the spatial domain, while channel attention scores feature data in the channel domain. Specifically, the module performs global average pooling on the feature data along the channel dimension to obtain W*H*1 data, where W*H is the spatial dimension of the input features. This W*H*1 data is then passed through a convolutional layer to score the importance of features in space. Next, a sigmoid function is used to limit the scoring range to (0,1), obtaining the weights of each spatial feature. Finally, these spatial weights are multiplied by the input features of the module to achieve attention to important spatial features. Figure 2 ).

[0039] c. Design a hybrid attention mechanism module, which is a combination of the aforementioned channel attention mechanism module and spatial attention mechanism module through a concat operation. Features from the convolutional combination layer are simultaneously input into both the channel attention mechanism module and the spatial attention mechanism module. The H*W*C dimensional features output by the channel attention module and the H*W*C dimensional features output by the spatial attention mechanism module are concatted to obtain an H*W*2C dimensional feature. This feature is then processed through convolution, batch normalization, and ReLU activation to restore it to H*W*C dimensionality, which is the output feature of this hybrid attention module. Figure 3 ).

[0040] d. Design a Unet neural network architecture incorporating a hybrid attention mechanism. The Unet neural network and the hybrid attention module are built using the PyTorch machine learning library. The Unet neural network includes 18 3×3 convolutional layers, 4 max-pooling downsampling layers, 4 deconvolutional upsampling layers, 1 1×1 convolutional layer, and 3 skip connections. The activation function of the convolutional layers is the ReLU function, and the activation function of the output layer is the Sigmoid function. By adding the hybrid attention mechanism module to the skip connection part of the regular Unet neural network, the hybrid attention mechanism can be integrated. Figure 4 ).

[0041] e. Design a loss function for the Unet neural network that incorporates a hybrid attention mechanism. This loss function J... SAUnet It can be represented as:

[0042] J SAUnet =J MSE +λJ Attention

[0043] J MSE Let J be the mean squared error loss function. Attention Let λ represent the loss function constraint term of the hybrid attention module, and λ represent the coefficient of the adjustment constraint term. MSE It can be represented as:

[0044]

[0045] Where N represents the total number of input data; x represents the label value; x' represents the input data; k represents the sequence number of the input data and the corresponding label, and its value range is all integers in the interval [1, N]. J Attention It can be represented as:

[0046]

[0047] Where m and n represent the maximum number of pixels in the horizontal and vertical directions of the feature map, respectively; F and F' represent the feature map input to the hybrid attention module and the predicted feature map of the hybrid attention module, respectively; i and j represent the horizontal and vertical pixel indices of the feature map, respectively, and their values ​​range from all integers in the intervals [1, m] and [1, n].

[0048] f. Classify the data from actual tunnel seismic wave method advanced geological prediction and establish training and test datasets. Extract complete data with high signal-to-noise ratio, and manually sample data with random, continuous, and uniform trace missing values ​​to generate training data. Use the corresponding complete data as labels to establish the training dataset Tr. set The missing track numbers in the actual data of missing tracks that have been excavated and verified for rock mass properties are padded with zeros to serve as test data (Te). set The set is used to verify the effectiveness of the data reconstruction results.

[0049] g. Initialize network parameters: set the learning rate to 0.001, the learning rate decay rate to 0.01, and the number of training samples input each time to 1000, for a total of 1000 training iterations.

[0050] h, Tr set Using a sliding time window W s T Slicing the data into pieces yields the corresponding training data slices. Where T represents the maximum number of tunnel seismic data, and s represents the step size of each movement of the sliding window, i.e., the sampling interval of the seismic data.

[0051] i. will Input the Unet neural network parameters with a hybrid attention mechanism, adjust the hyperparameters and data volume, and perform multiple training runs. Compare the output results with the labels using FK spectrum and single-channel waveform analysis to determine the accuracy of the data reconstruction results. Take the network model parameters with the smallest loss function and the highest data reconstruction accuracy as the final training result.

[0052] j. Te set The data reconstruction test was performed on the trained Unet neural network model with a hybrid attention mechanism. The velocity inversion and structural surface imaging were performed on the reconstructed complete data, and the accuracy was compared with the excavation verification results.

[0053] Example 1

[0054] This embodiment provides an application of a Unet neural network-based method for reconstructing tunnel seismic wave method advanced geological prediction data using actual tunnel seismic wave method data reconstruction, specifically as follows:

[0055] Table 1. Actual tunnel seismic wave method advanced geological prediction data

[0056] 24 4 0.2s 0.0001s 0~2000Hz

[0057] The tunnel's seismic wave method for advanced geological prediction data ( Figure 5a A total of 24 channels were sampled, with channels 6, 11, 13, and 14 missing and padded with zeros. The total sampling time was 0.2 s, the sampling interval was 0.0001 s, and the data bandwidth was 0–2000 Hz. Figure 5a The output obtained from the Unet neural network model with a hybrid attention mechanism, after the data input is trained, is as follows: Figure 5b As shown, the reconstructed tunnel seismic data accurately recovers the missing seismic traces that were filled in, and the effective reflection signals are continuous and clear in phase. Figure 5c and Figure 5d The comparison of the FK spectra of the labeled data and the reconstructed data shows that the energy of the reconstructed data is consistent with that of the labeled data, i.e., the complete data, without any energy loss or abnormal increase. Furthermore, the amplitude and phase differences between the single-channel signal waveforms of the labeled data and the reconstructed data are very small. This proves that the Unet neural network tunnel seismic wave method for advanced geological prediction data reconstruction based on the integration of a hybrid attention mechanism can provide high-quality complete data for tunnel seismic wave method advanced geological prediction.

[0058] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for reconstructing tunnel seismic data incorporating a hybrid attention Unet neural network, characterized in that, Includes the following steps: a. Design a channel attention mechanism module, which consists of three parts: squeezing, activation, and channel attention calculation. The squeezing part compresses the feature information output by the convolutional layer by performing global average pooling or max pooling on the features by channel. The activation part includes Conv1-Act1-Conv2-Act2, where Conv represents the convolutional layer and Act represents the activation layer. The channel attention calculation part multiplies the weight values ​​with the feature maps output by the convolutional layer to achieve weighted processing of the channels, thereby achieving attention to important features. b. Design a spatial attention mechanism module, which scores feature data in the spatial domain, and channel attention, which scores feature data in the channel domain. c. Design a hybrid attention mechanism module, which is composed of the channel attention mechanism module and the spatial attention mechanism module mentioned above, combined by a Concat operation. The features from the convolutional combination layer are simultaneously input into the channel attention mechanism module and the spatial attention mechanism module. The features output by the channel attention mechanism module and the features output by the spatial attention mechanism module are combined by a Concat operation, and then activated by convolution, batch normalization and ReLU function. The feature is restored to the same size as the input feature map, which is the output feature of this hybrid attention module. d. Design a Unet neural network architecture that incorporates a hybrid attention mechanism. The Unet neural network and the hybrid attention module are built using the PyTorch machine learning library. The Unet neural network includes convolutional layers, downsampling layers, upsampling layers, and skip connections. By adding the hybrid attention mechanism module to the skip connection part of the regular Unet neural network, the hybrid attention mechanism can be integrated. e. Design a loss function for the Unet neural network that incorporates a hybrid attention mechanism. This loss function J SAUnet It can be represented as: J SAUnet =J MSE +λJ Attention Among them, J MSE Let J be the mean squared error loss function. Attention λ represents the loss function constraint term of the hybrid attention module, and λ represents the coefficient of the adjustment constraint term; f. Classify the actual tunnel seismic wave method advanced geological prediction data, establish training and test datasets, extract complete data with high signal-to-noise ratio, and manually sample data with random trace missing, continuous trace missing, and uniform trace missing to generate training data. The corresponding complete data are used as labels to establish the training dataset Tr. set The missing track numbers in the actual data of missing tracks that have been excavated and verified for rock mass properties are padded with zeros to serve as test data (Te). set The set is used to verify the effectiveness of the data reconstruction results; g. Initialize network parameters, determine the learning rate, learning rate decay rate, number of input samples per cycle, and number of learning iterations; h, Tr set Using sliding window Slicing the data into pieces yields the corresponding training data slices. Where T represents the maximum number of tunnel seismic data, and s represents the step size of each movement of the sliding window, i.e., the sampling interval of the seismic data; i. will Input the Unet neural network parameters with a hybrid attention mechanism, adjust the hyperparameters and data volume, and perform multiple training sessions. Compare the output results with the labels using FK spectrum and single-channel waveform analysis to determine the accuracy of the data reconstruction results. Take the network model parameters with the minimum loss function and the highest data reconstruction accuracy as the final training result. j. Te set The data reconstruction test was performed on the trained Unet neural network model with a hybrid attention mechanism. The velocity inversion and structural surface imaging were performed on the reconstructed complete data, and the accuracy was compared with the excavation verification results.

2. The method for reconstructing tunnel seismic data incorporating a hybrid attention Unet neural network according to claim 1, characterized in that, Step b, specifically, involves: first, performing global average pooling on the feature data along the channel dimension; then, passing it through a convolutional layer to score the spatial importance of the features; next, using the Sigmoid function to limit the scoring range to (0,1) to obtain the weights of each spatial feature; finally, multiplying the above spatial weights by the input features of the module to achieve attention to important spatial features.

Images