CN112305591B - Tunnel advance geological prediction method, computer readable storage medium - Google Patents

Abstract

The invention discloses a tunnel advance geological prediction method and a computer-readable storage medium. The tunnel advance geological prediction method comprises: arranging a three-dimensional observation system on a tunnel face to realize three-dimensional seismic wave excitation and reception; The collected data is processed to obtain the location of the abnormal body. The data processing based on the deep learning algorithm includes: removing the bad sectors of the collected data based on the AlexNet convolutional neural network model algorithm, and filtering; And the direct wave velocity is estimated; the linear dynamic correction is performed on the original data with the bad track data removed; the isotravel profile stacking is performed on the data after linear dynamic correction; the reflective layer is picked up on the stacked data. The tunnel advance geological prediction method and the computer-readable storage medium can improve the detection efficiency and detection accuracy of abnormal bodies.

Description

Tunnel advance geological prediction method, computer readable storage medium

technical field

The present invention relates to the technical field of tunnel engineering, in particular to a method for advance geological prediction of tunnels and a computer-readable storage medium.

Background technique

The tunnel advance forecasting technology is a technical method to detect the geological environment and geological structure in front of the tunnel face before and during the construction of the tunnel. Among them, the tunnel advanced prediction technology based on seismic detection detects the surrounding rock information in front of the tunnel face according to the elastic wave field of the tunnel surrounding rock, with large detection depth and high detection accuracy, and is one of the main advanced prediction technologies. According to different detection methods, the existing tunnel advance prediction technology based on seismic detection can be roughly divided into two-dimensional detection technology and quasi-three-dimensional detection technology. The two-dimensional detection technology mostly uses the tunnel wall or face to excite seismic waves by means of explosives, electronic detonators, hammering, etc. The observation system with three-component geophones arranged in a linear arrangement on the tunnel wall collects reflected seismic waves for advance prediction, such as HSP and TSP technologies. Wait. The detection technology of the quasi-three-dimensional observation system mostly adopts the three-dimensional spatial arrangement of the detector around the tunnel wall, the observation system that stimulates the reflected wave on the tunnel wall or the face, and adopts seismic tomography, scattered wave superposition, Keshihoff depth migration, etc. Seismic data processing methods for imaging, such as TRT, TST, TSWD, etc.

During the process of realizing the present invention, the inventor found that the advanced prediction technology based on seismic detection often has the problems of false detection or missed detection, low detection efficiency, long data processing period, and multiple interpretation results.

Specifically, the inventor found that although the observation system of two-dimensional detection technologies such as HSP and TSP technology has a certain spatial distribution, due to the limitation of the tunnel space, the seismic wave excitation and reception angle is extremely small, and the imaging quality is poor, making it difficult to obtain accurate geological Exception body information. Compared with the two-dimensional observation technology, the quasi-three-dimensional detection technologies such as TRT, TST, TSWD and other quasi-three-dimensional detection technologies have improved the wave velocity analysis accuracy and the imaging accuracy of unfavorable geological bodies. Due to the influence of the sub-surface, it is difficult to arrange the shot points and detection points in all directions, and it is easy to lose some phase data, so it is difficult to realize the real three-dimensional detection; Due to the interference of strong non-reflection information such as converted waves and horizontal strata parallel to the tunnel axis, the signal-to-noise ratio is low; at the same time, the existing tunnel advance prediction seismic data is mostly processed manually, which has a long period and high cost, and the processing results are affected by the outside world. Factors and the professional knowledge of technicians have a great influence, the imaging has strong multi-solution, and the tunnel advance prediction method based on the manual method cannot intuitively compare the tunnel advance prediction seismic data of different tunnel construction locations.

The information disclosed in this Background section is only for enhancement of understanding of the general background of the invention and should not be taken as an acknowledgement or any form of suggestion that this information forms the prior art already known to a person of ordinary skill in the art.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a tunnel advance geological prediction method and a computer-readable storage medium, which can improve the detection efficiency and detection accuracy of abnormal bodies.

In order to achieve the above purpose, the present invention provides a method for advance geological forecasting of tunnels. The method for advance geological forecasting of tunnels includes: arranging a three-dimensional observation system on a tunnel face to realize three-dimensional seismic wave excitation and reception; The data collected by the observation system is processed to obtain the location of the abnormal body in the tunnel. Wherein, arranging a three-dimensional observation system on the tunnel face to realize three-dimensional seismic wave excitation and reception includes: arranging an observation system on the face of the tunnel; at the first tunnel construction position, the observation system excites seismic waves and receives reflected waves to collect data . Wherein, processing the data collected by the three-dimensional observation system based on the deep learning algorithm to obtain the position of the abnormal body in the tunnel includes: removing bad track data from the collected data based on the AlexNet convolutional neural network model algorithm, and performing Filter; identify the first arrival wave based on the U-Net convolutional neural network model algorithm, and estimate the speed of the direct wave; perform linear dynamic correction on the original data with the bad track data removed according to the first arrival wave identification result; The superimposed data is carried out with equal travel profile; the superimposed data is picked up by the reflection layer through the deep learning algorithm, so as to infer the area where the abnormal body of the first tunnel construction position is located.

In an embodiment of the present invention, the observation system as a whole is centrally symmetric.

In an embodiment of the present invention, the detectors of the observation system are arranged radially from the center to the surrounding, and the seismic wave generating devices of the observation system are also arranged radially from the center to the surrounding.

In an embodiment of the present invention, the removing bad sector data from the collected data based on the AlexNet convolutional neural network model algorithm includes: training the AlexNet convolutional neural network model; using the trained AlexNet The convolutional neural network model removes the bad sector data from the collected data.

In an embodiment of the present invention, the AlexNet convolutional neural network model takes a seismic wave field as a benchmark according to the resolution of seismic exploration, takes a quarter wavelet wavelength as the convolution kernel length, and eighth One wavelet wavelength is used as the pooling layer length.

其中，N为标签种类，pre(i)代表AlexNet对每一道数据的属于某一类标签的可能性预测值，F(x)代表每一道数据的所有标签预测值之和，F(x)_max代表所有参与预测数据的最大所有标签预测值之和。In an embodiment of the present invention, the use of the trained AlexNet convolutional neural network model to perform bad-sector elimination includes: judging data whose predicted data value is outside the range of the effective track prediction threshold f(x) as noise data , and culling, where,

Among them, N is the label type, pre(i) represents the probability prediction value of AlexNet belonging to a certain type of label for each data, F(x) represents the sum of all label predictions for each data, F(x) _max Represents the maximum sum of all label predictions for all participating prediction data.

In an embodiment of the present invention, the step of picking up the reflection layer on the superimposed data by using a deep learning algorithm includes: using a principal component analysis algorithm to identify a similar area recorded by multiple shots, where the similar area is presumed to be the area where the geological anomaly is located .

In an embodiment of the present invention, the method for tunnel advance geological prediction further includes: after completing the estimation of the abnormal body position of the first tunnel construction position, constructing a seismic signal correlation based on a U-Net convolutional neural network model algorithm A model is used to estimate abnormal bodies in other construction positions of the same tunnel according to the seismic signal correlation model.

In an embodiment of the present invention, the construction of the seismic signal correlation model based on the U-Net convolutional neural network model algorithm, and the abnormal body inference at other construction positions of the same tunnel according to the seismic signal correlation model includes: obtaining Seismic single shot data of one or more tunnel construction positions; classify the reflected wave event axis in the seismic single shot data of each tunnel construction position to obtain the label data of the reflected wave event axis of each tunnel construction position; Based on the U-Net convolutional neural network algorithm, the label data of the reflected wave event axis of each tunnel construction location and the seismic single shot data of each tunnel construction location are trained to establish a seismic signal correlation model; obtain the current The seismic single shot data of the tunnel construction location; the reflected wave event axis in the seismic single shot data of the current tunnel construction location is identified and located according to the seismic signal correlation model; according to the surface elevation and surrounding rock information of the construction area and The position of the abnormal body in front of the current tunnel construction position is estimated based on the identification and positioning results of the reflected wave event axis in the seismic single shot data of the current tunnel construction position.

Based on the same inventive concept, the present invention also provides a computer-readable storage medium for performing the following steps: removing bad sector data from the data collected by the observation system based on the AlexNet convolutional neural network model algorithm, and filtering ; Identify the first arrival wave based on the U-Net convolutional neural network model algorithm, and estimate the speed of the direct wave; perform linear dynamic correction on the original data with the bad track data removed according to the first arrival wave identification result; The data are superimposed on the time-travel profile; the superimposed data is picked up by the reflection layer through the deep learning algorithm, so as to infer the area where the abnormal body of the first tunnel construction location is located.

In an embodiment of the present invention, the computer-readable storage medium is further configured to perform the following steps: acquiring single-shot seismic data of one or more tunnel construction positions; The reflected wave event axis is classified to obtain the label data of the reflected wave event axis of each tunnel construction position; the label data of the reflected wave event axis of each tunnel construction position based on the U-Net convolutional neural network algorithm and The seismic single shot data of each tunnel construction position is trained to establish a seismic signal correlation model; the seismic single shot data of the current tunnel construction position is obtained; according to the seismic signal correlation model, the earthquake signal of the current tunnel construction position is analyzed. Identify and locate the reflected wave event axis in the single shot data; according to the surface elevation and surrounding rock information in the construction area and the reflected wave event axis identification and positioning results in the seismic single shot data of the current tunnel construction location, the current tunnel is identified and located. The position of the abnormal body in front of the construction site is estimated.

Compared with the prior art, according to the tunnel advanced geological prediction method and the computer-readable storage medium of the present invention, the seismic data of a single construction location of the tunnel is quickly identified based on the deep learning algorithm, and the dependence on the professional knowledge of the data processing personnel is low, and the processing is easy. The result is less disturbed by external influences, and can realize intelligent, fast and low-cost processing of large amounts of data. Preferably, in one embodiment, the propagation law of the tunnel seismic wave field is used, and a centrally symmetric observation system, especially a radial observation system, can be used, so that a large amount of reflected wave information with higher reliability and authenticity can be obtained. Little impact on data processing and interpretation. Preferably, in one embodiment, the reflected wave event axis formed by the same abnormal body is used in the single shot data collected at different detection positions, which shows that the travel time is different and the waveform is similar. After the detection result of the location, the subsequent construction locations can use the similar features of the waveform to establish a correlation model. Specifically, the correlation model of the data collected at different construction locations is constructed based on the deep learning algorithm, which can avoid manual identification of large amounts of data. Due to the problems of low efficiency and poor reliability of special waveforms, rapid and accurate identification of characteristic waveforms in a large amount of data is realized, and the invention can intuitively compare and analyze the seismic data of tunnel advance prediction in different tunnel construction positions, which is very intelligent.

Description of drawings

Fig. 1 is the step composition of the tunnel advance forecasting method according to an embodiment of the present invention;

Fig. 2 is the energy distribution diagram of the tunnel space seismic wave field when the tunnel face excites the source according to an embodiment of the present invention;

Fig. 3 shows a centrally symmetric observation system 3a, and a three-dimensional seismic wave field energy distribution diagram 3b of the observation system;

FIG. 4 shows a non-centrosymmetric observation system 4a and a three-dimensional seismic wave field energy distribution diagram 4b of the observation system;

5 shows a matrix orthogonal observation system 5a, an octagonal orthogonal observation system 5b, a radial observation system 5c, a non-orthogonal observation system 5d and an observation system parameter table 5e;

Fig. 6 shows the shot point of four types of three-dimensional area observation systems and the azimuth angle distribution diagram of the connection line of the detection point;

Fig. 7 shows the distribution diagram of the coverage times of four types of three-dimensional area observation systems;

Fig. 8 is the influence of different shot line numbers or shot line angle and shot spacing of the radial observation system according to an embodiment of the present invention on the uniformity of azimuth distribution;

Fig. 9 is the influence of different shot lines or shot line angle and shot spacing on the uniformity of coverage times distribution of the radial observation system according to an embodiment of the present invention;

Figure 10 shows the original data 10a of the three-dimensional reflection seismic signal, the single-shot recording spectrum curve 10b, the single-channel time spectrum 10c of the 4th track, and the single-track time spectrum 10d of the 5th track collected by the radial observation system;

Fig. 11 shows the curve 11a of the loss function of the AlexNet convolutional neural network model training process as a function of the number of iterations, and the effective channel prediction threshold curve 11b;

FIG. 12 shows the actual data 12a before the bad sectors are eliminated and the actual data 12b after the bad sectors are eliminated;

Figure 13 shows the curve 13a of the loss function value in the training process of the U-Net convolutional neural network with the number of iterations, the first-arrival wave identification data recorded by a single shot and the recognition result 13b, and the first-arrival recorded by another single shot. Wave identification data and identification results 13c;

Figure 14 shows data 14a before linear motion correction and data 14b after linear motion correction;

FIG. 15 is a two-dimensional horizontal slice diagram superimposed on iso-travel time sections according to an embodiment of the present invention;

16 is a schematic diagram of identifying similar regions of multiple shots recorded using a principal component analysis algorithm according to an embodiment of the present invention;

Fig. 17 shows the automatic identification result 17a and the vertical section 17b of the reflected wave identification result of the similar area of the deep learning-based multi-shot record;

18 is a section construction report diagram according to an embodiment of the present invention;

19 is a numerical simulation model diagram of different construction positions corresponding to four different times according to an embodiment of the present invention;

20 is a deep learning training sample and label data of a construction position corresponding to time T1 according to an embodiment of the present invention;

21 is a deep learning training sample and label data of a construction position corresponding to time T2 according to an embodiment of the present invention;

22 is a deep learning training sample and label data of a construction position corresponding to time T3 according to an embodiment of the present invention.

23 is a deep learning training sample and label data of a construction position corresponding to time T4 according to an embodiment of the present invention;

FIG. 24 is a schematic diagram of correction of a data processing result of a single acquisition according to an embodiment of the present invention;

25 is a simulation diagram of data collection in different construction stages of four tunnels according to an embodiment of the present invention;

Figure 26 is data collected from experiments according to an embodiment of the present invention;

FIG. 27 is a curve showing the variation of the loss function value with the number of iterations according to an embodiment of the present invention;

FIG. 28 is a recognition result of a step waveform according to an embodiment of the present invention.

Detailed ways

The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

Unless expressly stated otherwise, throughout the specification and claims, the term "comprising" or its conjugations such as "comprising" or "comprising" and the like will be understood to include the stated elements or components, and Other elements or other components are not excluded.

In order to overcome the problems of the prior art, the tunnel advanced geological prediction method provided by the following embodiments, combined with the propagation law of the tunnel seismic wave field and the energy distribution law, proposes a rapid design and arrangement method of the tunnel face observation system, so as to realize the rapid development of the geophone. Arrangement, in the case of reducing the impact on the tunnel construction process and the damage of the face, to obtain the reflection information of the abnormal body in front of the face with higher signal-to-noise ratio and higher reliability; multiple data collection during the tunnel construction process , based on deep learning technology, rapid and intelligent processing of seismic data and rapid and intelligent identification of abnormal bodies, based on a large amount of acquired seismic reflection data, and finally obtain more real spatial location information of abnormal bodies.

Fig. 1 is a tunnel advance forecasting method according to an embodiment of the present invention, the tunnel advance forecasting method includes: performing three-dimensional seismic wave signal excitation and reception based on an observation system on the tunnel face; The data is processed to obtain the location of the anomaly in the tunnel. Specifically, the method includes steps S1 to S8.

In step S1, a three-dimensional seismic observation system is set on the face.

When setting up the observation system, the inventors first conducted the following research.

Figure 2 shows the energy distribution of the tunnel space seismic wave field when the tunnel face excites the source, in which the x-axis is the direction of the horizontal extension of the tunnel, the z-axis is the direction perpendicular to the ground, and the tunnel represents the tunnel. It can be seen from Fig. 2 that the energy distribution of the spatial seismic wave field on both sides of the tunnel is axisymmetric with respect to the excitation point when the source is excited by the tunnel face. Therefore, in the process of the actual observation system layout, firstly, it is judged whether the observation system is reasonable by evaluating the uniformity of the energy distribution of the observation system to the tunnel space seismic wave field.

The inventor found that an observation system that satisfies the center symmetry can avoid the difference in the energy distribution of the seismic wave field in the tunnel space caused by the arrangement of the observation system.

Fig. 3 shows a centrally symmetric observation system 3a, the three-dimensional seismic wave field energy distribution of the observation system Fig. 3b, the plane seismic wave field energy contour of the observation system at Z=50 Fig. 3c, the observation system at y =50 plane seismic wavefield energy contour figure 3d. The y-axis is a direction parallel to the ground and perpendicular to the X-axis, and the observation system includes multiple detectors (as shown by the inverted triangles in the figure) and multiple seismic wave generating devices (as shown by the circles in the figure). From the comparison of the boxes in the contour diagram 3c and the contour diagram 3d, it can be shown that the energy distribution of the tunnel space seismic wave field of the center-symmetric observation system is very uniform.

Figure 4 shows a non-centrosymmetric observation system 4a, the three-dimensional seismic wave field energy distribution of the observation system Figure 4b, the observation system at Z=50 plane seismic wave field energy contour Figure 4c, the observation system at y =50 Planar seismic wavefield energy contour in Figure 4d. From the comparison of the boxes in the contour diagram 4c and the contour diagram 4d, it can be shown that the energy distribution of the tunnel space seismic wave field of the center-symmetric observation system is different.

It can be seen from Fig. 3 and Fig. 4 that the energy distribution of the tunnel space seismic wave field of the centrally symmetric observation system is more uniform. Therefore, preferably, in this embodiment, the observation system is arranged as a center-symmetric observation system, so that data with higher reliability and higher signal-to-noise ratio can be collected by utilizing the propagation law of the tunnel seismic wave field.

In addition, the inventor also found that when the coverage times of the observation system are more uniform and the azimuth distribution of the line connecting the shot point and the receiver point is more uniform, the signals received by the observation system are more in line with the propagation law of the tunnel seismic wave field, so the data collected by the observation system can be further improved. Signal-to-noise ratio and reliability of raw data.

Specifically, the inventors analyzed four different types of three-dimensional area observation systems. As shown in FIG. 5, the four different types of three-dimensional area observation systems are a matrix orthogonal observation system 5a, an octagonal orthogonal observation system 5b, a radial observation system 5c, and a non-orthogonal observation system 5d. In addition, FIG. 5 also shows the arrangement parameter table 5e of the four observation systems. FIG. 6 is a distribution diagram of the azimuth angle of the connection line between the shot points and the detection points of the four types of three-dimensional area observation systems. FIG. 7 is a distribution diagram of the coverage times of the four types of three-dimensional area observation systems.

From Figure 6 and Figure 7, it can be found that the coverage times of the radial observation system and the azimuth distribution of the line connecting the shot point and the receiver point are the most uniform.

It should also be noted that, in the actual exploration process, under the premise of meeting the exploration needs, for the radial observation system, in order to reduce the cost and layout time of the observation system, it is necessary to take into account the azimuth angle between the shot point and the receiver point. The distribution uniformity should also take into account the distribution uniformity of the coverage times, select the appropriate shot line (the line connecting the shot points passing through the center) and the signal receiving line (the line connecting the detection points passing through the center), and select the appropriate seismic wave generation line. Number of devices (shots) and detectors (spots). Figure 8 shows the effect of different number of shot lines or the included angle of shot lines and shot spacing on the uniformity of the azimuth distribution of the radial observation system. Figure 9 shows the effect of different shot lines or shot line angle and shot spacing on the uniformity of coverage times distribution in the radial observation system.

After the observation system is set up, in step S2, the observation system collects 3D seismic data: at time T1 at a certain construction position on the face, the observation system excites seismic waves, and the detector receives signals to realize 3D seismic data collection.

FIG. 10 shows the original data 10a of the three-dimensional reflection seismic signal, the single-shot recording spectrum curve 10b, the single-channel time spectrum 10c of the fourth track, and the single-track time spectrum 10d of the fifth track, respectively, which are collected by the radial observation system of the present embodiment. . The area of the first arrival wave and the area where the effective reflected wave may exist are marked in the single-channel time spectrum.

In step S3, the collected raw data is removed from bad sectors based on the AlexNet convolutional neural network algorithm. Specifically, the AlexNet convolutional neural network model is first trained, in which the AlexNet convolutional neural network model is based on a seismic wave field according to the resolution of seismic exploration, and a quarter wavelet wavelength is used as the convolution The core length, one-eighth of the wavelet wavelength is used as the length of the pooling layer. Then perform bad sector culling based on the trained model.

其中，N为标签种类(即有效数据种类)，pre(i)代表AlexNet对每一道数据的属于某一类标签(有效数据)的可能性预测值，F(x)代表每一道数据的所有标签预测值之和，F(x)_max代表所有参与预测数据的最大所有标签预测值之和。In order to better distinguish between valid tracks and bad tracks (empty track data and noise data) to improve the effect of bad track culling, preferably, in this embodiment, in bad track culling, a valid track prediction threshold f ( x), where

Among them, N is the label type (that is, the valid data type), pre(i) represents the probability prediction value of AlexNet belonging to a certain type of label (valid data) for each data, and F(x) represents all the labels of each data. The sum of predicted values, F(x) _max represents the sum of the largest predicted values of all labels of all participating prediction data.

FIG. 11 respectively shows a curve 11a of the loss function of the AlexNet convolutional neural network model training process of the present embodiment varying with the number of iterations, and a curve 11b of the effective channel prediction threshold value.

It can be seen from curve 11a that as the number of iterations increases, the value of the loss function gradually decreases. After the number of iterations is greater than 400, the value of the loss function tends to 0, indicating that the AlexNet convolutional neural network model tends to be stable here, and this ends train.

According to the training sample data, the upper limit and lower limit of the effective track record prediction value corresponding to different sample labels can be obtained. As shown in the curve 11b, the upper limit and the lower limit are used as the effective track prediction threshold. If the predicted data value is within this range, it is judged as Valid value, if outside this range, it is noise. Since the original seismic data without any processing is used to construct the bad track identification network, there is some noise influence in the predicted value. Therefore, preferably, in this embodiment, the effective track prediction threshold is adjusted to a certain extent according to the signal-to-noise ratio. To expand, the data processing in this preferred embodiment expands the range shown in 11b by 50%, and finally determines that the upper limit of valid data is 63.2, and the lower limit is -94.5. The 48 common receiver point shot sets are imported into the trained convolutional neural network for bad sector identification.

FIG. 12 respectively shows the actual data 12a before bad track removal and the actual data 12b after bad track removal.

The actual data 12a before bad track removal is 48 original single-track record data, the upper curve is the predicted value curve of different single-track records, the horizontal axis represents the shot set records of different common receiving points, and the vertical axis represents the AlexNet convolutional neural network algorithm based on The predicted value of bad track identification, the smaller the predicted value, the greater the possibility of bad track, and the less likely it is to be bad track. Comparing the actual data 12a after bad track removal and the actual data 12b before bad track removal, it can be seen that high frequency, sinusoidal noise and low frequency, strong amplitude noise are effectively removed.

In step S4, filter processing is performed on the data after the bad sectors are eliminated.

In step S5, the first arrival wave is identified based on the U-Net convolutional neural network algorithm, and the direct wave velocity is estimated. First, the U-Net convolutional neural network model is trained, and then the first arrival waves are identified based on the trained convolutional neural network model, and then the direct wave velocity is estimated.

Figure 13 shows the change curve 13a of the loss function value with the number of iterations in the training process of the U-Net convolutional neural network, the first-arrival wave identification data and the identification result 13b recorded by a single shot, and the initial wave recorded by another single shot. Arrival wave identification data and identification result 13c.

As shown in the curve 13a of the change of the value of the loss function with the number of iterations, as the number of iterations increases, the value of the loss function of the convolutional neural network gradually decreases and finally tends to remain unchanged, wherein the value of the loss function decreases when the number of iterations is 0 to 30 times. The speed is fast. The loss function value decreases slowly in the range of 30 to 150 times. After more than 170 times, the loss function value tends to 0, indicating that the U-Net network has become stable at this time. In this embodiment, the number of iterations is used. The network model when it is equal to 1350 is used as the final network for first arrival identification.

From the first-arrival identification data and identification results 13b recorded by a single shot and the first-arrival identification data and identification results 13c recorded by another single shot, it can be seen that the deep learning technology based on the U-Net algorithm has effectively identified the First arrival wave location. In addition, as shown in Fig. 13b, the direct wave velocity is 4000 m/s.

In step S6, linear motion correction is performed: linear motion correction is performed on the data after bad track elimination according to the first arrival wave position identification result. Figure 14 shows the data 14a before linear motion correction and the data 14b after linear motion correction. It can be seen that the linear motion correction results are very good.

In step S7, isotravel profile stacking is performed. Specifically, the apparent velocity of the first arrival wave is used as the initial velocity to scan the velocity. When the velocity on the grid node is correct, the superposition of multiple isotravel profiles at this point shows a strong energy value, otherwise it is a weak energy value .

FIG. 15 is a two-dimensional horizontal slice diagram superimposed on iso-travel time sections of the present embodiment. Observing its two-dimensional horizontal slices, it can be found that the region x=50-300 can be divided into three parts: 1, 2, and 3 according to the energy change of the reflection event axis.

In step S8, the reflection layer is picked up based on a deep learning algorithm, and specifically, a principal component analysis algorithm is used to identify similar areas recorded by multiple shots. This similar area is presumed to be the area where the geological anomaly is located. As shown in Fig. 16, the area enclosed by the rectangle is estimated to be the area where the abnormal body is located.

Fig. 17 shows the automatic identification results 17a of similar regions recorded by multiple shots based on deep learning and the vertical section 17b of the reflected wave identification results. Similar regions are represented by the phase reversal of the event axis (the region shown by the black box in the figure). The energy of the horizontal section at y=100m is stronger, while the energy of the vertical section at z=100m is relatively weak. This is mainly because the azimuth angle of the line connecting the shot point and the receiver point is mainly distributed in the z direction and in the y direction. less, making the energy distribution stronger in this direction. At the same time, the reflected wave is picked up in combination with the three regions in the two-dimensional horizontal slice of FIG. 16 , and the pickup result is shown in FIG. 17 . According to the position and shape of the event axis, it is divided into three parts: x=100-150, x=200-250, and x=300, and it is speculated that there are abnormal body structures in these three regions.

In order to verify the presumed result, Fig. 18 is a construction report drawing of the 2+966-3+311 section of this embodiment, and the triangle position in the figure is the position of this detection. It can be seen from the construction report that there is a sausage-shaped quartz dyke and a granite pegmatite dyke interspersed in the rock at a position of 60 meters in front of the face (3+074 in Figure 18), and its position is perpendicular to the reflection wave identification result. In 17b, the reflected wave anomaly at a distance of 55 meters from the face is consistent (as shown in the rectangular box 1 in Fig. 17b); in the surrounding rock at a distance of 260 meters from the face, there is a width of 0.1-0.5m, The 2-10m long huanggangite dikes intruded along the fractures are interspersed, and in spatial position, it coincides with the event axis of Fig. 17b at 250 meters in front of the face (as shown by the rectangular box 3 in Fig. 17b); At the same time, there is a reflected wave with strong energy between 60 meters and 250 meters away from the face of the tunnel. Comparing it with the event axis at a distance of 60 meters from the face of the tunnel (as shown by the rectangular box 2 in Figure 17b), it can be seen that , the phase change law is similar, and it can be seen from the follow-up construction report that in the pile number 2+966-3+240 section, the integrity and stability of the surrounding rock are poor, and the fold deformation is serious, so it is inferred that the reflected wave is as shown in the figure Multiples formed by events within rectangular frame 1 in 17b.

In addition, the tunnel is a dynamic construction process. During the whole construction process, the position of the detection target does not change, and the detection position changes with the construction progress. The inventor found that this characteristic makes the reflected wave events formed by the same anomaly appear to have different travel times and similar waveforms on the single-shot data collected at different detection positions. Therefore, in order to improve the estimation efficiency of abnormal bodies in the future construction process, in a preferred embodiment, a U-Net convolutional neural network algorithm is used to construct a correlation model of seismic signals collected at different construction positions of the tunnel, and a computer These large amounts of data are subjected to intelligent learning of characteristic waves to realize intelligent identification and extraction of the reflection information of the same abnormal body in different seismic signals, thereby eliminating the influence of external factors such as ground cover, surface layer, and tunnel wall on the identification of abnormal bodies in front of the face. .

Specifically, after completing the estimation of the abnormal body position at the first moment, the process of estimating the abnormal body at each subsequent construction position is as follows: first obtain the seismic single shot data of one or more tunnel construction positions; The reflected wave event axis in the seismic single shot data is classified to obtain the label data of the reflected wave event axis of each tunnel construction location; based on the U-Net convolutional neural network algorithm, the reflected wave of each tunnel construction location The label data of the event axis and the seismic single shot data of each tunnel construction position are trained to establish a seismic signal correlation model; the seismic single shot data of the current tunnel construction position is obtained; Identify and locate the reflected wave event axis in the seismic single shot data at the current tunnel construction location; identify and locate the reflected wave event axis in the seismic single shot data at the current tunnel construction location based on the surface elevation and surrounding rock information in the construction area and the current tunnel construction location As a result, the position of the abnormal body in front of the current tunnel construction position is estimated.

It should be noted that the more sample data for training, the more accurate the prediction result of abnormal body. Therefore, after estimating the position of the abnormal body at the first construction position, according to the above process, the subsequent position of the abnormal body at the second construction position is obtained; then when estimating the position of the abnormal body at the third construction position, preferably, the The single shot data and label data of the first construction position and the second construction position are used as training samples for training; and so on, when inferring the abnormal body position of the fourth construction position, preferably, the first construction position should be , the second construction position and the single shot data and label data of the third construction position are used as training samples for training. In this way, a more accurate abnormal body estimation result can be obtained.

Next, the intelligent interpretation technology of tunnel advance forecast data is explained with numerical simulation data. Figure 19 is a numerical simulation model of different construction positions corresponding to four different times. FIG. 20 shows the deep learning training samples and label data of the construction location corresponding to time T1. FIG. 21 shows the deep learning training samples and label data of the construction position corresponding to time T2. FIG. 22 shows the deep learning training samples and label data of the construction position corresponding to time T3. FIG. 23 shows the deep learning training samples and label data of the construction position corresponding to time T4. In order to increase the randomness of the sampling data, the distances between the four construction moments are not equidistant; The tunnel face corresponding to time T3 is located at x=53m, and the tunnel face corresponding to time T4 is located at x=73m. A total of 20 data tracks were recorded for each single shot, and the track spacing was 0.5m. Firstly, an initial multi-time single-shot data event correlation numerical simulation is established by using the single-shot data at T1 time. Observing the data at the time of T1, it can be seen that there are two types of reflected wave events in the reflected wave area: the negative apparent velocity event axis (at 75ms) and the event axis with small apparent inclination angle (at 50ms). The event axis is divided into two categories. The event axis at 50ms and small apparent inclination angle is regarded as the first type event axis, and the negative apparent velocity event axis at 75ms is regarded as the second type event axis, so as to make the label data at T1 time (as shown in the figure). 20). The shot collection records show that as the tunnel gets closer to the position of the anomaly, the event axis of the reflected wave formed by the anomaly gradually becomes smaller when traveling, and the depth of the horizontal stratum changes little in the range of x=25 and x=75. The reflected wave formed by the bulk influence travels almost unchanged. Next, based on the deep learning technology, the event axis of the shot set recording at different tunnel construction positions is automatically identified, as shown in Figures 21-23. It can be seen that the multi-time seismic data correlation network model established based on the deep learning algorithm can effectively identify and reflect the change law of the reflected wave event axis formed by the same abnormal body. Secondly, due to the high similarity of the waveforms of multiples and primarys, when only the T1 data is used as training data to build a multi-time shot record correlation model, the prediction results are poor, and it is difficult to separate the primary and multiples. As a result, a large amount of multiple wave information remains (as shown in Figure 21); however, when the training sample data increases, the multi-time shot record correlation model constructed based on multiple data can not only effectively distinguish primary waves from multiples (such as As shown in Figure 22 and Figure 23), the anti-interference ability of waveform recognition has also been enhanced. For example, in the shot collection record at the time of T4, the reflected wave of the effective abnormal body has been aliased with the direct wave. When the data at three times is used as the When training data, the recognition ability of the event axis is significantly improved, which indicates that the prediction result is positively correlated with the training data set. On this basis, the data processing results of a single acquisition are corrected. As shown in Figure 24, the interference of the ground surface and the stratum parallel to the tunnel axis has been effectively filtered out.

In addition, in order to verify the effect of eigenwave identification by establishing a correlation model, a description will be given using a physical experiment. First, two survey lines were arranged using a smartsolo 5Hz nodal seismometer, each with 20 receiving points and a track spacing of 0.5m. The process of pedestrian walking along the survey line is similar to the process of tunnel excavation forward. When walking to a certain receiving point, it can be considered that the tunnel construction has reached a certain position, and the physical experiment is used to simulate the tunnel excavation process. During the data collection process, the survey line was walked back and forth twice, and the data collection at different construction stages of the tunnel was simulated four times, as shown in Figure 25. Take the footsteps recorded by the pedestrian for the first time as the reflected wave received by the tunnel at time T1 (shown in the m box in Figure 26), and take the footsteps recorded in the following three times as the reflected wave signal collected at different subsequent times (the n box in Figure 26). shown). In order to increase the data difference, passers-by held a 5kg lead put during the first walk, and two 5kg lead put in the next three walks. The data in the m frame is used as the training data, and the track number is used as the sample label for training. On this basis, the data in the n frame is identified. The change curve of the loss function value with the number of iterations is shown in Figure 27. The recognition results shown in Figure 28 show that the neural network model constructed using the data at time T1 can effectively identify the position of the subsequent footstep waveforms. This experiment shows that this method can be used to establish the correlation model of the single shot data collected at different times in the tunnel, so that the This enables eigenwave identification.

Based on the same inventive concept, this embodiment also provides a computer-readable storage medium for performing the following steps: removing bad sector data from the data collected by the observation system based on the AlexNet convolutional neural network model algorithm, and performing the following steps: Filter; identify the first arrival wave based on the U-Net convolutional neural network model algorithm, and estimate the speed of the direct wave; perform linear dynamic correction on the original data with the bad track data removed according to the first arrival wave identification result; The superimposed data is carried out with equal travel profile; the superimposed data is picked up by the reflection layer through the deep learning algorithm, so as to infer the area where the abnormal body of the first tunnel construction position is located.

Preferably, the computer-readable storage medium of this embodiment is further configured to perform the following steps: acquiring single-shot seismic data of one or more tunnel construction positions; Classify to obtain the label data of the reflected wave event axis of each tunnel construction position; based on the U-Net convolutional neural network algorithm, the label data of the reflected wave event axis of each tunnel construction position and the each The seismic single shot data of the tunnel construction location is trained to establish a seismic signal correlation model; the seismic single shot data of the current tunnel construction location is obtained; according to the seismic signal correlation model, the seismic single shot data of the current tunnel construction location is analyzed. Identify and locate the reflected wave event axis; according to the surface elevation and the surrounding rock information of the construction area and the reflected wave event axis identification and positioning results in the seismic single shot data of the current tunnel construction location The location of the abnormal body is estimated.

To sum up, according to the tunnel advanced geological prediction method and the computer-readable storage medium of the present embodiment, the seismic data of a single construction location of the tunnel is quickly identified based on the deep learning algorithm, and the dependence on the professional knowledge of the data processing personnel is low, and the processing results are affected by the outside world. The impact and interference are small, and the intelligent, fast and low-cost processing of large amounts of data can be realized. Preferably, the seismic observation system on the tunnel face is proposed for the first time by using the propagation law of the seismic wave field in the tunnel. The observation system of center symmetry is selected, especially the radial observation system. It is verified from theory and practice that the observation system can obtain A large number of more reliable and more realistic reflected wave information, and at the same time have little impact on data processing and interpretation. Preferably, in one embodiment, the reflected wave event axis formed by the same abnormal body is used in the single shot data collected at different detection positions, which shows that the travel time is different and the waveforms are similar, and only the detection result of one position needs to be obtained. After that, the subsequent construction positions can use the similar characteristics of the waveform to establish the correlation model. Specifically, the correlation model of the data collected at different construction positions can be constructed based on the deep learning algorithm, which can avoid the low efficiency of manual identification of special waveforms in a large amount of data. , the problem of poor reliability, to achieve rapid and accurate identification of characteristic waveforms in a large amount of data.

As will be appreciated by those skilled in the art, the embodiments of the present application may be provided as a method, a system, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may 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.) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It will be understood that each process and/or block in the flowchart illustrations and/or block diagrams, and combinations of processes and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions The apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.

These computer program instructions can also be loaded on a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that The instructions provide steps for implementing the functions specified in the flow or blocks of the flowcharts and/or the block or blocks of the block diagrams.

The foregoing descriptions of specific exemplary embodiments of the present invention have been presented for purposes of illustration and illustration. These descriptions are not intended to limit the invention to the precise form disclosed, and obviously many changes and variations are possible in light of the above teachings. The exemplary embodiments were chosen and described for the purpose of explaining certain principles of the invention and their practical applications, to thereby enable others skilled in the art to make and utilize various exemplary embodiments and various different aspects of the invention. Choose and change. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims (6)

1. A method for advance geological prediction of a tunnel is characterized by comprising the following steps:

exciting and receiving three-dimensional seismic wave signals based on an observation system on a tunnel face; and

processing the data acquired by the observation system based on a deep learning algorithm to obtain the position of the abnormal body in the tunnel;

the three-dimensional seismic wave signal excitation and reception based on the observation system on the tunnel face comprises the following steps:

arranging a three-dimensional earthquake observation system on the tunnel face;

at a first tunnel construction position, the three-dimensional earthquake observation system excites earthquake waves and receives reflected waves so as to acquire data;

wherein the processing the data collected by the observation system based on the deep learning algorithm to obtain the position of the abnormal body in the tunnel comprises:

based on AlexNet convolutional neural network model algorithm, removing bad channel data from the collected data, and filtering;

identifying the first-motion wave based on a U-Net convolution neural network model algorithm, and estimating the speed of the direct wave;

performing linear dynamic correction on the original data without the bad track data according to the first arrival wave identification result;

carrying out equal travel time section superposition on the data after linear dynamic correction; and

carrying out reflection layer pickup on the superposed data through a deep learning algorithm, thereby inferring the area of the abnormal body at the first tunnel construction position;

the method for removing the bad track data from the collected data based on the AlexNet convolutional neural network model algorithm comprises the following steps:

training the AlexNet convolutional neural network model;

removing bad track data from the acquired data by adopting the trained AlexNet convolutional neural network model;

wherein, the step of adopting the trained AlexNet convolutional neural network model to remove the bad channel comprises the following steps:

judging the data with the data prediction value outside the range of the effective channel prediction threshold f (x) as noise data and removing the noise data,

，

，

wherein, N is label type, pre (i) represents the possibility predicted value of AlexNet belonging to a certain type label for each data, F (x) represents the sum of all label predicted values of each data, F (x)_maxRepresenting the sum of the maximum all-label predictors of all participating predictors;

wherein the step of performing reflector picking on the superimposed data through a deep learning algorithm comprises:

identifying a similar region of the multi-shot records by adopting a principal component analysis algorithm, wherein the similar region is presumed to be a region where the geological abnormal body is located;

the method for forecasting the advance geology of the tunnel further comprises the following steps:

after the estimation of the position of the abnormal body at the first tunnel construction position is completed, a seismic signal correlation model is built based on a U-Net convolution neural network model algorithm, and the estimation of the abnormal body is carried out at other construction positions of the same tunnel according to the seismic signal correlation model.

2. The method for advanced geological forecasting of tunnels as claimed in claim 1, wherein the observation system is centrally symmetric as a whole.

3. A method for advanced geological prediction of a tunnel as claimed in claim 2, wherein the geophones of said observation system are arranged radially from the center to the periphery, and the seismic generators of said observation system are also arranged radially from the center to the periphery.

4. The method for look-ahead geological prediction of a tunnel of claim 1, wherein the AlexNet convolutional neural network model is based on a seismic wavelet field according to seismic exploration resolution, with one-quarter of a wavelet wavelength as the length of the convolutional kernel and one-eighth of a wavelet wavelength as the length of the pooling layer.

5. The method for forecasting tunnel geology according to claim 1, wherein the building of the seismic signal correlation model based on the U-Net convolutional neural network model algorithm, and the estimation of the abnormal body at other construction positions of the same tunnel according to the seismic signal correlation model comprises the following steps:

acquiring seismic single shot data of one or more tunnel construction positions;

classifying the reflected wave homophase axes in the seismic single shot data of each tunnel construction position to obtain the label data of the reflected wave homophase axes of each tunnel construction position;

training the label data of the reflection wave homophase axis of each tunnel construction position and the seismic single shot data of each tunnel construction position based on a U-Net convolution neural network algorithm, and establishing a seismic signal correlation model;

acquiring seismic single shot data of a current tunnel construction position;

identifying and positioning a reflected wave event in the seismic single shot data of the current tunnel construction position according to the seismic signal correlation model;

and according to the earth surface elevation, the surrounding rock information of the construction area and the reflected wave homophase axis identification and positioning result in the seismic single-shot data of the current tunnel construction position, the abnormal body position in front of the current tunnel construction position is presumed.

6. A computer-readable storage medium based on the method for advanced geological prediction of tunnels according to any of claims 1 to 5, characterized in that it is adapted to perform the following steps:

based on AlexNet convolutional neural network model algorithm, removing bad channel data from data collected by an observation system, and filtering;

identifying the first-motion wave based on a U-Net convolution neural network model algorithm, and estimating the speed of the direct wave;

performing linear dynamic correction on the original data without the bad track data according to the first arrival wave identification result;

carrying out equal travel time section superposition on the data after linear dynamic correction; and

carrying out reflection layer pickup on the superposed data through a deep learning algorithm, thereby inferring the area of the abnormal body at the first tunnel construction position;

the method for removing the bad track data from the collected data based on the AlexNet convolutional neural network model algorithm comprises the following steps:

training the AlexNet convolutional neural network model;

removing bad track data from the acquired data by adopting the trained AlexNet convolutional neural network model;

wherein, the step of adopting the trained AlexNet convolutional neural network model to remove the bad channel comprises the following steps:

judging the data with the data prediction value outside the range of the effective channel prediction threshold f (x) as noise data and removing the noise data,

，

，

wherein, N is label type, pre (i) represents the possibility predicted value of AlexNet belonging to a certain type label for each data, F (x) represents the sum of all label predicted values of each data, F (x)_maxRepresenting the sum of the maximum all-label predictors of all participating predictors;

wherein the step of performing reflector picking on the superimposed data through a deep learning algorithm comprises:

identifying a similar region of the multi-shot records by adopting a principal component analysis algorithm, wherein the similar region is presumed to be a region where the geological abnormal body is located;

the method for forecasting the advance geology of the tunnel further comprises the following steps:

after the estimation of the position of the abnormal body at the first tunnel construction position is completed, a seismic signal correlation model is built based on a U-Net convolution neural network model algorithm, and the estimation of the abnormal body is carried out at other construction positions of the same tunnel according to the seismic signal correlation model.

Images