TBM tunneling parameter prediction method based on passive source seismic wave signals

Abstract

The invention relates to a TBM (Tunnel Boring Machine) tunneling parameter prediction method based on a passive source seismic wave signal, which comprises the following steps of: arranging a detector to collect a passive source seismic wave signal propagated through surrounding rock, and rearranging the detector to collect the passive source seismic wave signal after the TBM is tunneled for a specified distance every time; converting the collected passive source seismic wave signal into a virtual source seismic record, converting a time domain seismic record into a distance domain seismic record, and extracting a reflected wave signal; extracting instantaneous characteristic parameters of the reflected wave signals, and jointly constructing an input vector matrix corresponding to similar tunneling parameters of an unexcavated area of the TBM in combination with historical tunneling parameters of the TBM; constructing a hybrid neural network model based on CNN and Autoformer, and training to determine an optimal hybrid neural network model; and taking the extracted instantaneous characteristic parameters of the seismic wave signals collected in real time and the historical tunneling parameters of the TBM as the input of the network model at the same time, and predicting the optimal tunneling parameters of the specified distance in front of the tunneling face of the TBM. Accurate prediction of the TBM tunneling parameters of the long-distance to-be-excavated area of the tunnel is achieved, normal construction of the TBM is not affected, and natural matching is achieved for TBM parameter prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of TBM tunneling parameter prediction, and particularly relates to a TBM tunneling parameter prediction method based on passive source seismic wave signals. Background Technique

[0002] TBM (Tunnel Boring Machine) has been widely used in tunnel excavation. The control of TBM tunneling parameters is crucial for the safe construction of tunnels. Currently, the adjustment of some main tunneling parameters (such as thrust value, torque value, cutterhead rotation speed value, tunneling speed) is mainly based on manual experience and the geological exploration data provided in the early stage. However, due to the subjectivity of manual experience and the uncertainty of geological conditions, it is impossible to provide a reliable and effective basis for the adjustment of tunneling parameters. Therefore, the adjustment of tunneling parameters usually cannot meet the construction needs, and it is easy to occur TBM jamming events in poor geological sections, resulting in construction period delays and increased construction costs.

[0003] During the TBM construction process, predicting the possible geological risks in front of the tunnel in advance has become an essential part of tunnel construction. According to the prediction results, the performance of the tunneling machine is evaluated and related tunneling parameters are adjusted in a timely and effective manner to achieve the purpose of reducing or eliminating geological disasters during the construction period and ensuring production safety.

[0004] Patent CN115711667A discloses a method for predicting TBM tunneling parameters based on vibration signals. During the TBM tunneling process, the vibration signals of the TBM are monitored, the main tunneling parameters of the TBM and the vibration data corresponding to the time are extracted. The tunneling parameters and vibration signals are split into training sets, validation sets, and test sets according to the same time period. The vibration signals are processed, and wavelet transform is performed on the vibration data to obtain time-domain diagrams and frequency-domain diagrams. A deep learning model of tunneling parameters and vibration time-frequency diagrams is established, and the model is trained through the training set data. Through the test set data, the error of the established model is optimized. Through the validation set data, the model indexes are evaluated. The vibration signals are input into the model to predict the tunneling parameters.

[0005] In the above patent, on the one hand, during the TBM tunneling process, the vibration signals of the TBM are monitored, but the geological changes that can be reflected by the vibration signals of the TBM are very limited. On the other hand, due to the training, testing, and validation of the model, the input is the time-domain diagram after wavelet transform of the vibration signal, and the vibration signal is input into the model to predict the tunneling parameters. Therefore, the prediction of TBM tunneling parameters in this patent is the prediction of the tunneling parameters at the current moment, and the input of the prediction model does not include the relevant characteristic parameters at different distances in front of the TBM tunneling face, so it is impossible to predict the tunneling parameters of the unexcavated area of the TBM at a long distance, and the prediction accuracy is also relatively low.

[0006] Therefore, how to improve the prediction method of TBM tunneling parameters in the prior art, enhance the prediction accuracy and reliability of TBM tunneling parameters, perform real-time dynamic adjustment, effectively improve the tunneling efficiency of TBM, and ensure the safe tunneling of the tunnel are technical problems that urgently need to be solved at present. Summary of the Invention

[0007] The object of the present invention is to provide a method for predicting TBM tunneling parameters based on passive source seismic wave signals, so as to improve the prediction method of TBM tunneling parameters in the prior art, enhance the prediction accuracy and reliability of TBM tunneling parameters, perform real-time dynamic adjustment, effectively improve the tunneling efficiency of TBM, and ensure the safe tunneling of the tunnel.

[0008] To solve the above technical problems, the technical solution adopted by the present invention is as follows: A method for predicting TBM tunneling parameters based on passive source seismic wave signals includes the following steps: S1: Arrange a plurality of geophones at specified positions on the tunnel sidewall. When the TBM is tunneling normally, collect the passive source seismic wave signals propagated through the surrounding rock in real time. After the TBM tunnels a specified distance, rearrange a plurality of geophones at the specified positions on the sidewall to collect the passive source seismic wave signals propagated through the surrounding rock; S2: Convert the collected passive source seismic wave signals propagated through the surrounding rock into virtual source seismic records, convert the time-domain seismic records into distance-domain seismic records, and extract the reflected wave signals; S3: Extract the instantaneous characteristic parameters of the reflected wave signals, and jointly construct an input vector matrix corresponding to the same type of tunneling parameters in the unexcavated area of the TBM in combination with the historical tunneling parameters of the TBM; S4: Construct a hybrid neural network model based on CNN and Autoformer, obtain the optimal hyperparameter combination of the model to train the model, and determine the optimal hybrid neural network model through multi-fold cross-validation; S5: Use the instantaneous characteristic parameters of the real-time collected seismic wave signals and the historical tunneling parameters of the TBM as the inputs of the network model at the same time to predict the optimal tunneling parameters at a specified distance in front of the TBM tunneling face.

[0009] Preferably, the specific process of step S1 is as follows: S11: Drill holes to a depth of 30 cm or more into the surrounding rock, insert coupling rods, and connect the geophones to the coupling rods; S12: When the TBM is tunneling normally, the cutterhead breaking rock will generate seismic wave signals, and use the arranged geophones to collect the passive source seismic wave signals propagated through the surrounding rock; S13: After the TBM tunnels a specified distance, iteratively execute steps S12 - S13.

[0010] Preferably, the number of deployed geophones is greater than or equal to 6, and the spacing is kept consistent during the deployment of geophones. The time for collecting the passive source seismic wave signals propagated through the surrounding rock is not less than 15 minutes.

[0011] In order to collect the seismic wave signals propagated through the surrounding rock, it is necessary to drill the hole depth more than 30 cm into the surrounding rock. Then, the geophone is fixed by the geophone coupling rod. Each data collection is carried out when the TBM is in a fully tunneling state. In order to ensure the effect of wave field characteristic restoration, long-term observation is required to obtain sufficient passive source data. Therefore, all geophones need to collect continuous vibration signals for not less than 15 minutes.

[0012] Preferably, the specific process of extracting the reflected wave signal in step S2 is as follows: S21: Perform seismic interference processing and superposition on the passive source seismic wave signals in segments to obtain a seismic record with the first geophone point as the virtual source and the other points as the receiving points; S22: Extract the arrival time of the direct wave from the first geophone to the last geophone, combine the arrival time of the direct wave and the distance, calculate the direct wave velocity, and convert the time-domain seismic record into a distance-domain seismic record based on this velocity; S23: Use the τ-p transform to filter out the direct wave and surface wave components in the distance-domain seismic record to extract the reflected wave signal in the seismic record.

[0013] For the virtual source seismic record obtained after processing the original signal and the extracted reflected wave signal, since the seismic record contains various types of signals, such as direct waves, surface waves, reflected waves, transmitted waves, etc., and the reflected wave is mainly generated by the change of the medium wave impedance, it can best reflect the change of the geological situation. Therefore, the τ-p transform algorithm is used to extract the reflected wave signal.

[0014] Preferably, the specific formula for extracting the arrival time of the direct wave from the first geophone to the last geophone in step S22, combining the arrival time of the direct wave and the distance, and calculating the direct wave velocity is: ; ; In the formula , s i , t i are the direct wave velocities of the i th ( i >1) geophone with the first geophone as the source, the distance between the i th ( i >1) geophone and the first geophone, and the distance between the i th ( i> 1) The arrival time of the direct wave of the geophone. v p is the finally calculated direct wave velocity, which is 3700 m / s. Based on this wave velocity, the time-domain seismic record is converted into a distance-domain seismic record.

[0015] Preferably, the specific process of extracting the instantaneous characteristic parameters of the reflected wave signal in step S3 and jointly constructing the input vector matrix corresponding to the same type of tunneling parameters in the unexcavated area of the TBM in combination with the historical tunneling parameters of the TBM is as follows: S31: Perform empirical mode function decomposition on the reflected wave signal to achieve frequency-domain decomposition of the signal and effective separation of each component. In the separation result, filter out the ultra-high frequency and very low frequency components. The ultra-high frequency is the component with a main frequency greater than 10000 Hz, and the very low frequency is the component with a main frequency less than 20 Hz; S32: Perform Hilbert transform on the remaining components, calculate the instantaneous amplitude and instantaneous frequency at different distance positions of each component, and perform the same data processing on all seismic signals collected each time to obtain the instantaneous characteristic parameter matrix of the corresponding signals; S33: Record the tunneling parameters of the TBM when tunneling through different positions of the tunnel, extract the tunneling parameters within a specified distance in the already excavated area behind the predicted position, and jointly construct the input vector matrix with the instantaneous characteristic parameter matrix, and output the tunneling parameter value at a specified distance in front of the TBM tunneling face.

[0016] The instantaneous amplitude is a measure of the reflection intensity, which can reflect the propagation and attenuation characteristics of seismic waves in different geological bodies. The change of the instantaneous amplitude can indicate the change of formation lithology, the location of the aquifer, and possible geological anomalies such as faults or cavities, and can also reflect the characteristics of the underground medium boundary. The instantaneous frequency is a measure of the phase time change rate, which is not affected by the energy size, and can reflect the change of the lithology of the formation. When seismic waves pass through some special formations, the high-frequency components will be strongly attenuated, which can reflect the change of the formation medium composition and possible geological structures such as faults and fracture zones, and can well identify the medium boundary. Therefore, the instantaneous amplitude and instantaneous frequency are selected as the characteristic quantities reflecting geological changes, which have a large correlation with the tunneling parameters. In addition, due to the certain correlation of the time sequence before and after the tunneling parameters, the prediction accuracy of short distances can be improved. Therefore, the tunneling parameters within a specified distance range in front of the predicted position are also used as the input characteristic quantities for predicting the same type of tunneling parameters of the subsequent TBM. The selection range of the historical tunneling parameters of the TBM is mainly selected in combination with the length of the predicted distance. The longer the predicted distance, the smaller the selection range of the historical tunneling parameters of the TBM.

[0017] Preferably, in step S4, a hybrid neural network model based on CNN and Autoformer is constructed, the optimal hyperparameter combination of the model is obtained to train the model, and the specific process of determining the optimal hybrid neural network model through multi-fold cross-validation is as follows: S41: The established hybrid neural network model is composed of a CNN network and an Autoformer network. The CNN network is composed of two one-dimensional CNN layers and a max-pooling layer stacked, and a dropout layer is added; the Autoformer network is composed of an encoder stack layer and a decoder stack layer; S42: The CNN network model and the Autoformer network model are merged through a Concatenate layer, and a dropout layer is added to prevent overfitting, and finally the output is through a fully connected layer; S43: The instantaneous feature parameter matrix in step S32 and the time series of tunneling parameters within a specified distance in the already excavated range behind the predicted position are respectively used as the inputs of the established hybrid CNN-Autoformer prediction model, and the output is set as the tunneling parameter value at a specified position in front of the TBM tunneling face; S44: The Bayesian optimization algorithm is used to optimize the number of nodes in the CNN hidden layer, the number of encoder stack layers of the Autoformer, the number of decoder stack layers, and the best initial learning rate in the model to obtain the optimal hyperparameter combination; S45: The model is trained based on the optimal hyperparameter combination, and the corresponding optimal hybrid neural network model is selected through multi-fold cross-validation.

[0018] The beneficial effects of the present invention include: The TBM tunneling parameter prediction method based on passive source seismic wave signals provided by the present invention arranges geophones to collect passive source seismic wave signals propagated through the surrounding rock, and re-arranges geophones to collect passive source seismic wave signals after the TBM tunnels each specified distance; the collected passive source seismic wave signals are converted into virtual source seismic records, the time-domain seismic records are converted into distance-domain seismic records, and the reflected wave signals are extracted; the instantaneous feature parameters of the reflected wave signals are extracted, and combined with the TBM historical tunneling parameters, an input vector matrix corresponding to predicting the same type of tunneling parameters in the unexcavated area of the TBM is jointly constructed; a hybrid neural network model based on CNN and Autoformer is constructed and trained to determine the optimal hybrid neural network model; the instantaneous feature parameters of the real-time collected seismic wave signals and the TBM historical tunneling parameters are used as the inputs of the network model at the same time to predict the optimal tunneling parameters at a specified distance in front of the TBM tunneling face.

[0019] Since the vibration signal generated during the normal rock breaking of the TBM contains rich information related to the surrounding rock geological conditions after being propagated in the surrounding rock, by collecting this passive source seismic wave signal, extracting the instantaneous characteristic parameters of the reflected wave signal therein, and combining with the historical tunneling parameters of the TBM, they are jointly used as the input for predicting the same type of tunneling parameters in the unexcavated area of the TBM. By constructing a hybrid deep learning neural network model, the prediction of the optimal tunneling parameters of the TBM is realized, which can effectively improve the tunneling efficiency of the TBM and ensure the safe tunneling of the tunnel. The present invention realizes the accurate prediction of the TBM tunneling parameters in the long-distance unexcavated area of the tunnel, and the collection of the passive source seismic wave signal has no impact on the normal construction of the TBM, and has a natural matching for the parameter prediction of the TBM, and can provide a crucial safety guarantee for the future realization of intelligent tunneling of the TBM. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic flow chart of the TBM tunneling parameter prediction method based on passive source seismic wave signals of the present invention.

[0021] Figure 2 (a) is a schematic diagram of the geophone layout.

[0022] Figure 2 (b) is a photo of the on-site layout of the geophone of the present invention.

[0023] Figure 3 It is a schematic diagram of the virtual source seismic record obtained after processing the original signal by the present invention and the main reflected wave signals extracted therefrom.

[0024] Figure 3 (a) is the virtual source seismic record; Figure 3 (b) is the extracted reflected wave signal.

[0025] Figure 4 It is the instantaneous amplitude and instantaneous frequency of each component at different distance positions after filtering by the present invention.

[0026] Figure 4 (a) is the instantaneous amplitude of component 1 at different positions.

[0027] Figure 4 (b) is the instantaneous frequency of component 1 at different positions.

[0028] Figure 4 (c) is the instantaneous amplitude of component 2 at different positions.

[0029] Figure 4 (d) is the instantaneous frequency of component 2 at different positions.

[0030] Figure 4 (e) is the instantaneous amplitude of component 3 at different positions.

[0031] Figure 4 (f) Instantaneous frequencies of component 3 at different positions.

[0032] Figure 5 is the average thrust value of each ring of the TBM thrust recorded in the present invention at different positions.

[0033] Figure 6 is a schematic diagram of the CNN-Autoformer hybrid neural network model architecture of the present invention.

[0034] Figure 7 is a comparison chart of the prediction results obtained by the optimal hybrid neural network model of the present invention on the training set data and the best thrust value adopted during actual tunneling.

[0035] Figure 7 (a) Prediction results of the training set for the 10th ring in front of the TBM tunneling face.

[0036] Figure 7 (b) Prediction results of the training set for the 20th ring in front of the TBM tunneling face.

[0037] Figure 8 is a comparison chart of the average thrust values predicted for the 10th and 20th rings in front of the TBM tunneling face by the present invention using the newly extracted data and the best thrust value adopted during actual tunneling.

[0038] Figure 8 (a) Prediction results of the validation set for the 10th ring in front of the TBM tunneling face.

[0039] Figure 8 (b) Prediction results of the validation set for the 20th ring in front of the TBM tunneling face. Detailed implementation manner

[0040] The following further describes the present invention in detail with reference to the appended Figures 1 to 8 drawings: Example 1 Referring to the appended Figure 1 drawings, a method for predicting TBM tunneling parameters based on passive source seismic wave signals includes the following steps: S1: A plurality of geophones are arranged at designated positions on the tunnel sidewall. When the TBM is tunneling normally, passive source seismic wave signals propagated through the surrounding rock are collected in real time. After the TBM tunnels a designated distance, a plurality of geophones are re-arranged at designated positions on the sidewall to collect passive source seismic wave signals propagated through the surrounding rock; S2: The passive source seismic wave signals propagated through the surrounding rock collected are converted into virtual source seismic records, the time-domain seismic records are converted into distance-domain seismic records, and the reflected wave signals are extracted; S3: Extract the instantaneous characteristic parameters of the reflected wave signal, and jointly construct an input vector matrix corresponding to the same type of tunneling parameters in the unexcavated area of the TBM by combining with the historical tunneling parameters of the TBM; S4: Construct a hybrid neural network model based on CNN and Autoformer, obtain the optimal hyperparameter combination of the model to train the model, and determine the optimal hybrid neural network model through multi-fold cross-validation; S5: Use the instantaneous characteristic parameters of the real-time collected seismic wave signal and the historical tunneling parameters of the TBM as the input of the network model at the same time to predict the optimal tunneling parameters at a specified distance in front of the TBM tunneling face. The historical tunneling parameters of the TBM in step S3 and the optimal tunneling parameters at different distances in front of the TBM tunneling face predicted in S5 both include the thrust value, torque value, cutter head rotation speed value, and tunneling speed.

[0041] By performing modal decomposition and time-frequency analysis on the seismic wave signal generated during TBM tunneling, relevant instantaneous characteristic parameters are extracted, providing reliable seismic wave data support for the geological conditions at different positions of the tunnel. At the same time, combined with the historical tunneling thrust value of the TBM, they are jointly used as the input for predicting the tunneling thrust in the unexcavated area of the TBM. By constructing a hybrid neural network model based on CNN and Autoformer, reliable prediction of the TBM tunneling thrust value is achieved for real-time dynamic adjustment, effectively improving the tunneling efficiency of the TBM and ensuring the safe tunneling of the tunnel. Compared with the existing methods, on the one hand, the present invention learns a large amount of seismic wave data based on the deep learning algorithm, and by establishing the relationship between the instantaneous characteristics of seismic signals and the TBM tunneling thrust, reduces the multi-solution problem of geophysical exploration inversion methods, and realizes the long-distance prediction of the TBM tunneling parameters in the unexcavated area of the tunnel. On the other hand, the passive source seismic wave signal acquisition used has no impact on the normal construction of the TBM, has a natural match for the parameter prediction of the TBM, and can provide crucial safety guarantees for the future realization of intelligent TBM tunneling.

[0042] Example 2 On the basis of Example 1, the specific process of step S1 is as follows: S11: Use the grouting holes on the tunnel support segment to drill into the surrounding rock more than 30 cm deep, insert the coupling rod, and connect the geophone to the coupling rod; S12: When the TBM is tunneling normally, the cutter head breaking the rock will generate a seismic wave signal. Use the arranged geophones to collect this passive source seismic wave signal propagated through the surrounding rock; S13: After the TBM advances every 50 m, iteratively execute steps S12 - S13.

[0043] The number of geophones deployed is greater than or equal to 6, and the spacing is kept consistent during the deployment of geophones. The time for collecting the passive source seismic wave signals propagated through the surrounding rock is not less than 15 minutes.

[0044] In order to collect the seismic wave signals propagated through the surrounding rock, it is necessary to drill the hole depth more than 30 cm into the surrounding rock. Then, the geophones are fixed using the geophone coupling rod. Each data collection is carried out when the TBM is in a fully tunneling state. In order to ensure the effect of wave field characteristic restoration, long-term observation is required to obtain sufficient passive source data. Therefore, all geophones need to collect continuous vibration signals for not less than 15 minutes.

[0045] See Figure 2 (a), Figure 2 (b) for the on-site layout photos of geophones. In the double-shield TBM construction tunnel, a total of 6 geophones are laid on the left wall on site, and the spacing between adjacent geophones is 3.6 m. Since the tunnel is supported by segments with a length of 1.8 m, in order to collect the seismic wave signals propagated through the surrounding rock, first, drill holes using the grouting holes, and the hole depth is drilled more than 30 cm into the surrounding rock. Then, the geophones are fixed using the coupling rod and based on a strong magnet. The seismic wave signals are collected once every 50 m of TBM tunneling. Each collection is carried out when the TBM is in a fully tunneling state. In order to ensure the effect of wave field characteristic restoration, long-term observation is required to obtain sufficient passive source data. Therefore, all geophones collect continuous vibration signals for not less than 15 minutes, and a total of 91 data are collected. At the same time, the TBM tunneling thrust corresponding to each ring within the tunneling range is recorded, and the average value of the thrust values of each ring is taken, totaling 2,530 rings with a total length of 4,554 m.

[0046] Example 3 Based on Example 1 or Example 2, the specific process of extracting the reflected wave signals in step S2 is as follows: S21: Perform seismic interference processing and superposition on the passive source seismic wave signals in segments to obtain a seismic record with the first geophone point as the virtual source and the remaining points as the receiving points; S22: Extract the arrival time of the direct wave from the first geophone to the last geophone. Combine the arrival time of the direct wave and the distance to calculate the direct wave velocity. Based on this velocity, convert the time-domain seismic record to a distance-domain seismic record. The specific formula for calculating the direct wave velocity is: ; ; In the formula , s i , t i are respectively the source with the first geophone, the i th (i > 1) The direct wave velocity of the geophone, the i th ( i > 1) The distance between the geophone and the first geophone, the i th ( i > 1) The arrival time of the direct wave of the geophone. v p is the finally calculated direct wave velocity, which is 3700 m / s. Based on this wave velocity, the time-domain seismic record is converted into a distance-domain seismic record.

[0047] S23: Use the τ-p transform to filter out the direct wave and surface wave components in the distance-domain seismic record to extract the reflected wave signal in the seismic record.

[0048] The virtual source seismic record obtained after processing the original signal and the extracted reflected wave signal. Since the seismic record contains various types of signals, such as direct waves, surface waves, reflected waves, transmitted waves, etc., and the reflected wave is mainly generated by the change of the medium wave impedance, it can best reflect the change of the geological situation. Therefore, the τ-p transform algorithm is used to extract the reflected wave signal.

[0049] Example 4 Based on Example 1 or Example 2 or Example 3, refer to Figure 3 (a), Figure 3 (b) The virtual source seismic record obtained after processing the original signal, and the extracted main reflected wave signal. The specific process of extracting the instantaneous characteristic parameters of the reflected wave signal in step S3 and jointly constructing the input vector matrix corresponding to the same type of tunneling parameters in the unexcavated area of the TBM is as follows: S31: Perform empirical mode function decomposition on the reflected wave signal to achieve frequency-domain decomposition of the signal and effective separation of each component. In the separation result, filter out the ultra-high frequency and very low frequency components. The ultra-high frequency is the component with a main frequency greater than 10000 Hz, and the very low frequency is the component with a main frequency less than 20 Hz; S32: Perform Hilbert transform on the remaining components, calculate the instantaneous amplitude and instantaneous frequency at different distance positions of each component, and perform the same data processing on all seismic signals collected each time to obtain the instantaneous characteristic parameter matrix of the corresponding signal; S33: Record the tunneling parameters when the TBM tunnels through different positions of the tunnel, extract the tunneling parameters at a specified distance within the excavated area behind the predicted position, and jointly construct the input vector matrix with the instantaneous characteristic parameter matrix, and the output is set as the tunneling parameter value at a specified distance in front of the TBM tunneling face.

[0050] Instantaneous amplitude is a measure of reflection intensity, which can reflect the propagation and attenuation characteristics of seismic waves in different geological bodies. The change of instantaneous amplitude can indicate the change of formation lithology, the location of aquifers, and geological anomalies such as possible faults or cavities, and can also reflect the characteristics of the underground medium boundary. Instantaneous frequency is a measure of the phase time change rate, which is not affected by the energy size and can reflect the change of lithology composing the formation. When seismic waves pass through some special formations, the high-frequency components will be strongly attenuated, which can reflect the change of formation medium composition and possible geological structures such as faults and fracture zones, and can well identify the medium boundary. Therefore, instantaneous amplitude and instantaneous frequency are selected as characteristic quantities reflecting geological changes, which have a large correlation with tunneling parameters. In addition, due to the certain correlation of the time series before and after tunneling parameters, the accuracy of short-distance prediction can be improved. Therefore, the tunneling parameters within a specified distance range before the prediction position are also used as input characteristic quantities for predicting the same type of tunneling parameters of the subsequent TBM. The selection range of TBM historical tunneling parameters is mainly selected in combination with the length of the predicted distance. The longer the predicted distance, the smaller the selection range of TBM historical tunneling parameters. See Figure 4 (a), Figure 4 (b), Figure 4 (c), Figure 4 (d), Figure 4 (e), Figure 4 (f), the instantaneous amplitude and instantaneous frequency of each component at different distance positions after filtering. Since there is a certain correlation in the time series before and after tunneling parameters, which can improve the accuracy of short-distance prediction, the thrust value within a specified distance range before the prediction position is also used as an input characteristic quantity for predicting the subsequent TBM tunneling thrust value. See Figure 5 , the average thrust value of each ring of the recorded TBM thrust at different positions, with a total of 2530 rings and a total length of 4554m.

[0051] Example 5 On the basis of Example 1 or Example 2 or Example 3 or Example 4, in step S4, a hybrid neural network model based on CNN and Autoformer is constructed, the optimal hyperparameter combination of the model is obtained to train the model, and the specific process of determining the optimal hybrid neural network model through multi-fold cross-validation is as follows: S41: The established hybrid neural network model is composed of a CNN network and an Autoformer network. The CNN network is composed of two one-dimensional CNN layers and a max-pooling layer stacked, and a dropout layer is added; the Autoformer network is composed of an encoder stack layer and a decoder stack layer; S42: The CNN network model and the Autoformer network model are merged through the Concatenate layer, a dropout layer is added to prevent overfitting, and finally the output is through a fully connected layer; S43: Use the instantaneous feature parameter matrix in step S32 and the time series of tunneling parameters within a specified distance in the already excavated area behind the predicted position as the inputs of the established hybrid CNN-Autoformer prediction model, and set the output to the tunneling parameter values at a specified position in front of the TBM tunneling face. S44: Use the Bayesian optimization algorithm to optimize the number of nodes in the CNN hidden layer, the number of encoder stacks and decoder stacks in Autoformer, and the optimal initial learning rate in the model to obtain the optimal hyperparameter combination. S45: Train the model based on the optimal hyperparameter combination and select the corresponding optimal hybrid neural network model through multi-fold cross-validation.

[0052] See Figure 7 (a)、 Figure 7 (a) and (b), the inputs of the CNN network model are the instantaneous feature amplitudes and instantaneous frequencies of the reflected wave signals at different positions. First, use min-max normalization to preprocess the input data and adjust it to the standard scale (between 0 and 1) to obtain the input of the convolutional layer, with the shape of (B, T, 2), where B is the input batch size, T is the time step, and 2 represents the two features of instantaneous amplitude and instantaneous frequency. After entering the convolutional layer, use multiple convolutional kernels to extract local features. The size of the convolutional kernel can be adjusted according to the characteristics of the data. Usually, a smaller convolutional kernel (such as 3x3) is selected to capture local relationships. Activation function: Use the ReLU activation function to introduce non-linearity. Then enter the max-pooling layer to reduce the dimension of the data, reduce the computational amount, and add a dropout layer with a dropout probability of 0.2 to reduce overfitting. Finally, use the flatten layer to flatten the output of the convolutional layer into a one-dimensional vector to obtain the output shape of (B, T, CNN_output_size), where B is the batch size, T is the time step, and CNN_output_size is the feature quantity extracted by the CNN network for merging with the output of Autoformer.

[0053] The input of the Autoformer network model is the excavated multi-ring TBM tunneling thrust time series, that is, 1D time series data with the shape of (B, T, 1), where B is the batch size and T is the time step. This data is input into a multi-layer encoder stack layer, and each stack layer contains an auto-correlation mechanism and a feed-forward network to calculate the auto-correlation of the time series, discover the periodic patterns in the time series, and perform non-linear transformation on the features. The input of the decoder stack layer consists of: the output of the encoder stack layer and the trend part decomposed from the time series. After being processed by the auto-correlation mechanism and the feed-forward network in the decoder stack layer, the time series feature representation after multi-layer decoding is obtained, and these features contain the prediction information of future time steps. The output of this Autoformer is a time series with the shape of (B, T, 1), representing the prediction of the TBM tunneling thrust.

[0054] By processing through the above two networks respectively, the local features of multi-variables (instantaneous amplitude and instantaneous frequency of seismic wave signals) and the long-term dependence relationship of the TBM thrust time series can be effectively extracted. Then, the Concatenate layer is used to merge the output of the CNN part and the output of the Autoformer decoder in the feature dimension, and the merged shape is (B, T, CNN_output_size + 1). Then it enters the dropout layer with a dropout probability of 0.25 to prevent overfitting. Finally, the fully connected layer is used to map the merged feature vector to the final output dimension, and a scalar value, that is, the predicted TBM tunneling thrust value, is finally output.

[0055] Based on Bayesian optimization, the optimal hyperparameters are obtained for model training, and the mean squared error (MSE) is used as the loss function to measure the difference between the predicted value and the true value. The corresponding optimal hybrid neural network model is selected through 10-fold cross-validation. See Figure 7 The comparison chart of the prediction results obtained by the optimal hybrid neural network model on the training set data and the best thrust value used during actual tunneling. Extract the instantaneous feature parameters of the newly collected seismic wave signals and use them as the input of the network model together with the TBM historical tunneling parameters to predict the optimal tunneling parameters at different distances in front of the TBM tunneling face.

[0056] Figure 8 (a), Figure 8Figure (b) shows the comparison results between the predicted average thrust values of the 10th ring (18 m) and the 20th ring (36 m) in front of the TBM tunneling face using the newly extracted data and the optimal thrust values adopted during actual tunneling. From the comparison results, it can be seen that the mean absolute percentage error values (MAPE) reach 6.27% and 7.34% respectively, and the root mean square error values (RMSE) are 1177.19 and 1417.85 respectively. The method proposed by the present invention can accurately predict the optimal thrust value of the unexcavated area in front of the TBM tunneling face. Since the independent variables used to predict the TBM tunneling thrust value include the instantaneous characteristic parameters of the seismic signal indicating the change of the geological conditions in front of the tunnel, it can provide reliable data support for the prediction of the long-distance thrust value. At the same time, the tunneling thrust of the excavated section of the TBM is also used as an independent variable, which improves the prediction result of the short-distance thrust value.

[0057] In summary, for the TBM tunneling parameter prediction method based on the passive source seismic wave signal provided by the present invention, since the vibration signal generated during the normal rock-breaking tunneling of the TBM contains rich information related to the surrounding rock geological conditions after propagating through the surrounding rock, by collecting this passive source seismic wave signal and extracting the instantaneous characteristic parameters of the reflected wave signal, and at the same time combining the historical tunneling parameters of the TBM, they are jointly used as the input for predicting the same type of tunneling parameters in the unexcavated area of the TBM. By constructing a hybrid deep learning neural network model, the prediction of the optimal tunneling parameters of the TBM is realized, which can effectively improve the tunneling efficiency of the TBM and ensure the safe tunneling of the tunnel. The present invention realizes the accurate prediction of the TBM tunneling parameters in the long-distance unexcavated area of the tunnel, and the collection of the passive source seismic wave signal has no impact on the normal construction of the TBM. It has a natural match for the parameter prediction of the TBM and can provide a crucial safety guarantee for the future realization of intelligent tunneling of the TBM.

Claims

1. A method for predicting TBM excavation parameters based on passive source seismic wave signals, characterized in that: The following steps are involved: S1: Multiple geophones are arranged at designated positions on the tunnel side walls. When the TBM is advancing normally, passive source seismic wave signals propagated through the surrounding rock are collected in real time. After each TBM advances a designated distance, multiple geophones are re-arranged at designated positions on the side walls to collect passive source seismic wave signals propagated through the surrounding rock. S2: Convert the acquired passive source seismic wave signals after propagation through the surrounding rock into virtual source seismic records, convert the time domain seismic records into distance domain seismic records, and extract the reflection wave signals; S3: Extract the instantaneous characteristic parameters of the reflected wave signal, combine them with the historical excavation parameters of the TBM, and jointly construct the input vector matrix corresponding to the similar excavation parameters of the TBM unexcavated area; S4: Build a hybrid neural network model based on CNN and Autoformer, obtain the optimal hyperparameter combination of the model to train the model, and determine the optimal hybrid neural network model through multi-fold cross validation; S5: The extracted instantaneous characteristic parameters of the real-time collected seismic wave signals and the historical excavation parameters of the TBM are simultaneously used as the input of the network model to predict the optimal excavation parameters at a specified distance in front of the TBM excavation face.

2. The method for predicting TBM excavation parameters based on passive source seismic wave signals according to claim 1, characterized in that: The specific process of step S1 is as follows: S11: Drill a hole to 30 cm or more inside the surrounding rock, insert a coupling rod, and connect the detector to the coupling rod; S12: When the TBM is excavating normally, the cutter head will break the rock and generate seismic wave signals. The passive source seismic wave signals after propagation through the surrounding rock are collected by the deployed geophones. S13: After the TBM advances a specified distance each time, steps S12-S13 are iteratively executed.

3. The method for predicting TBM excavation parameters based on passive source seismic wave signals according to claim 2, characterized in that: The number of deployed geophones is greater than or equal to 6, and the time interval between the deployment of geophones is kept consistent, and the time for collecting passive source seismic wave signals after propagation through the surrounding rock is not less than 15 minutes.

4. The method for predicting TBM excavation parameters based on passive source seismic wave signals according to claim 1, characterized in that: The specific process of extracting the reflected wave signal in step S2 is as follows: S21: Perform seismic interference processing on the passive source seismic wave signal in segments and superimpose them to obtain a seismic record with the first detection point as a virtual seismic source and the remaining points as receiving points; S22: extracting the arrival time of the direct wave from the first geophone to the last geophone, combining the arrival time and distance of the direct wave, calculating the direct wave velocity, and converting the time domain seismic record into the distance domain seismic record based on the velocity; S23: Use τ-p transformation to filter out the direct wave and surface wave components in the distance domain seismic record to extract the reflected wave signal in the seismic record.

5. The method for predicting TBM excavation parameters based on passive source seismic wave signals according to claim 4, characterized in that: In step S22, the arrival time of the direct wave from the first detector to the last detector is extracted, and the specific formula for calculating the direct wave velocity is as follows, combining the direct wave arrival time and distance: ； ； In the formula , s i , t i The first geophone is used as the source, the second i indivual( i >1) The direct wave velocity of the detector, i indivual( i >1) The distance between the detector and the first detector, the i indivual( i >1) Arrival time of the direct wave from the detector; v p is the final calculated direct wave speed, which is 3700m / s; The time domain seismic record is converted into the distance domain seismic record based on the wave velocity.

6. The method for predicting TBM excavation parameters based on passive source seismic wave signals according to claim 1, characterized in that: The historical excavation parameters of the TBM in step S3 and the optimal excavation parameters at different distances in front of the TBM excavation face predicted in step S5 both include thrust value, torque value, cutter head speed value and excavation speed.

7. The method for predicting TBM excavation parameters based on passive source seismic wave signals according to claim 1, characterized in that: The specific process of extracting the instantaneous characteristic parameters of the reflected wave signal in step S3 and combining them with the historical excavation parameters of the TBM to jointly construct the input vector matrix corresponding to the similar excavation parameters of the unexcavated area of ​​the TBM is as follows: S31: performing intrinsic mode function decomposition on the reflected wave signal to achieve frequency domain decomposition of the signal and effective separation of each component, and filtering out ultra-high frequency and very low frequency components in the separation result, wherein the ultra-high frequency is a component with a main frequency greater than 10000 Hz, and the very low frequency is a component with a main frequency less than 20 Hz; S32: performing Hilbert transform on the remaining components, calculating the instantaneous amplitude and instantaneous frequency of each component at different distances, performing the same data processing on all seismic signals collected each time, and obtaining an instantaneous characteristic parameter matrix of the corresponding signal; S33: Record the excavation parameters of the TBM when it excavates through different positions of the tunnel, extract the excavation parameters of a specified distance behind the predicted position within the excavated range, construct an input vector matrix together with the instantaneous characteristic parameter matrix, and output the excavation parameter value set as the specified distance in front of the TBM excavation face.

8. The method for predicting TBM excavation parameters based on passive source seismic wave signals according to claim 1, characterized in that: In step S4, a hybrid neural network model based on CNN and Autoformer is constructed, the optimal hyperparameter combination of the model is obtained to train the model, and the specific process of determining the optimal hybrid neural network model through multi-fold cross validation is as follows: S41: The established hybrid neural network model consists of a CNN network and an Autoformer network, where the CNN network consists of two layers of one-dimensional CNN layers and a maximum pooling layer stacked, and a random inactivation layer is added; the Autoformer network consists of an encoder stack layer and a decoder stack layer; S42: The CNN network model and the Autoformer network model are merged through the Concatenate layer, a random inactivation layer is added to prevent overfitting, and finally the output is through the fully connected layer; S43: using the instantaneous characteristic parameter matrix in step S32 and the tunneling parameter time series of the specified distance in the excavated range behind the prediction position as the input of the established hybrid CNN-Autoformer prediction model, and setting the output as the tunneling parameter value of the specified position in front of the TBM tunneling face; S44: The Bayesian optimization algorithm is used to optimize the number of CNN hidden layer nodes, the number of Autoformer encoder stack layers, the number of decoder stack layers and the optimal initial learning rate in the model to obtain the optimal hyperparameter combination; S45: Train the model based on the optimal hyperparameter combination, and select the corresponding optimal hybrid neural network model through multi-fold cross validation.

Images