Real-time monitoring method and system for tunnel soft rock deformation

Abstract

The invention discloses a tunnel soft rock deformation real-time monitoring method and system. The method comprises the steps that firstly, a strain signal, displacement data, an AE signal and deformation data of tunnel soft rock are obtained through corresponding monitoring equipment arranged in a tunnel; then, corresponding time sequence features are extracted from the strain signals, the displacement data, the AE signals and the deformation data respectively and are aligned and combined into a multi-source time sequence feature vector; and finally, inputting the multi-source time sequence feature vectors into a trained LSTM prediction model, and performing fusion analysis on four-dimensional feature vectors in the multi-source time sequence feature vectors through the LSTM prediction model, thereby accurately distinguishing rock mass deformation caused by structural fracture and rock mass deformation caused by mechanical vibration, and accurately predicting the displacement of rock mass deformation, so as to improve the rock mass deformation accuracy. And high-precision prediction of deformation caused by rock mass structure fracture is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the deformation field for metering solid, and particularly relates to a tunnel soft rock deformation real-time monitoring method and system. BACKGROUND

[0002] In tunnel engineering, soft rock (such as mudstone, shale, fault fracture zone, etc.) is extremely prone to cause large deformation due to its unique rheological property, strain softening and swelling, which threatens the construction safety. The traditional monitoring method (such as single-point displacement meter, etc.) is limited by low sampling frequency (<1Hz), low spatial resolution (>1m) and poor anti-interference ability, and is difficult to meet the needs of deep tunnel. Although distributed optical fiber, micro inertial measurement and other technologies provide new ideas, the existing schemes lack multi-source data collaboration capability. For example, the existing technology of strain combined with displacement can capture deformation, but cannot distinguish rock mass rupture from mechanical vibration. The existing technology of strain combined with acoustic emission can identify rupture, but lacks displacement quantification capability. Therefore, it is difficult to accurately predict the deformation caused by the structure rupture of soft rock. SUMMARY

[0003] In view of the deficiencies in the prior art, the present application provides a tunnel soft rock deformation real-time monitoring method and system, which can accurately predict the rock mass deformation caused by the structure rupture of tunnel soft rock. The specific technical solutions are as follows: In a first aspect, a tunnel soft rock deformation real-time monitoring method is provided. In a first implementation manner of the first aspect, the method comprises: obtaining strain signals, displacement data, AE signals and deformation data of the tunnel soft rock; performing feature extraction on the strain signals, displacement data, AE signals and deformation data respectively to obtain corresponding time sequence features to form a multi-source time sequence feature vector; predicting the deformation caused by the structure rupture of the tunnel soft rock by an LSTM prediction model according to the multi-source time sequence feature vector.

[0004] In a second implementation manner of the first aspect, in combination with the first implementation manner of the first aspect, the feature extraction on the strain signals comprises: performing differential signal processing on the strain signals, and performing feature extraction on the processed strain signals.

[0005] In a third implementation manner of the first aspect, in combination with the first implementation manner of the first aspect, the feature extraction on the strain signals comprises: performing feature extraction on the strain signals by using a convolutional neural network.

[0006] In a fourth implementation manner of the first aspect, in combination with the first implementation manner of the first aspect, the feature extraction on the displacement data comprises: The displacement data is notch filtered, and the processed displacement data is subjected to feature extraction.

[0007] In combination with the first implementation manner of the first aspect, in the fifth implementation manner of the first aspect, the feature extraction on the displacement data comprises: The node displacement data of all nodes in the displacement data is fused by using a federal Kalman filtering algorithm.

[0008] In combination with the first implementation manner of the first aspect, in the sixth implementation manner of the first aspect, the feature extraction on the AE signal comprises: The AE signal is subjected to feature extraction by using a wavelet transform.

[0009] In combination with the first implementation manner of the first aspect, in the seventh implementation manner of the first aspect, the deformation amount of the tunnel soft rock is predicted by using an LSTM prediction model, comprising: The contribution weight of each time sequence feature in the multi-source time sequence feature vector is adjusted according to the lithology of the tunnel soft rock.

[0010] In combination with the first implementation manner of the first aspect, in the eighth implementation manner of the first aspect, further comprising: matching the predicted deformation amount with a multi-level early warning triggering condition, and triggering a corresponding early warning signal.

[0011] The second aspect provides a tunnel soft rock deformation real-time monitoring system, comprising: A data acquisition module configured to acquire strain signals, displacement data, AE signals and deformation data of a tunnel soft rock; A feature extraction module configured to respectively extract features from the strain signals, displacement data, AE signals and deformation data, and obtain corresponding time sequence features to form a multi-source time sequence feature vector; A deformation prediction module configured to predict a deformation amount of the tunnel soft rock caused by structural rupture of rock mass according to the multi-source time sequence feature vector by using an LSTM prediction model.

[0012] Advantages: The tunnel soft rock deformation real-time monitoring method and system can cooperatively use strain signals, displacement data, AE signals and deformation data to predict a deformation amount of a tunnel soft rock caused by structural rupture. The AE signals collected by an AE sensor and the deformation data of the tunnel soft rock collected by a laser radar are coupled and analyzed to distinguish between soft rock deformation caused by structural rupture and non-structural noise. The coupled analysis result is verified by combining strain signals and displacement data to improve the prediction result accuracy. The deformation data can correct temperature drift errors of the strain signals in reverse to further improve the prediction result accuracy, and realize high-precision prediction of the deformation amount of the tunnel soft rock. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the specific embodiments of the present application, the drawings required to be used in the specific embodiments will be briefly introduced below. In all the drawings, the elements or parts are not necessarily drawn according to the actual proportion.

[0014] Figure 1 The flow chart of the tunnel soft rock deformation real-time monitoring method provided by an embodiment of the present application. Specific embodiments

[0015] The embodiments of the technical solutions of the present application will be described in detail below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and therefore only serve as examples, but cannot limit the protection scope of the present application.

[0016] As shown in the flow chart of the tunnel soft rock deformation real-time monitoring method, the monitoring method comprises: Figure 1 Step 1, obtaining strain signals, displacement data, AE signals and deformation data of the tunnel soft rock; Step 2, respectively extracting features from the strain signals, displacement data, AE signals and deformation data to obtain corresponding time sequence features to form a multi-source time sequence feature vector; Step 3, predicting the deformation amount of the tunnel soft rock caused by the structural rupture of the rock mass according to the multi-source time sequence feature vector through an LSTM prediction model. Specifically, first, the strain signals, displacement data, AE signals and deformation data of the tunnel soft rock can be obtained through the corresponding monitoring equipment arranged in the tunnel, such as monitoring the strain signals of the tunnel soft rock through a distributed optical fiber vibration sensor, monitoring the AE signals through an acoustic emission sensor, monitoring the displacement data of the tunnel soft rock through a micro inertial measurement unit, and scanning the deformation data of the tunnel soft rock through a laser radar.

[0017] Then, strain time sequence features, displacement time sequence features, AE signal time sequence features and deformation time sequence features can be extracted from the strain signals, displacement data, AE signals and deformation data respectively. The strain time sequence features, displacement time sequence features, AE signal time sequence features and deformation time sequence features are aligned and combined into a multi-source time sequence feature vector.

[0018] Finally, the multi-source time sequence feature vector can be input into the trained LSTM prediction model, and the four-dimensional feature vectors in the multi-source time sequence feature vector can be fused and analyzed through the LSTM prediction model, so as to accurately distinguish the rock mass deformation caused by structural rupture and the rock mass deformation caused by mechanical vibration, and accurately predict the displacement amount of rock mass deformation, thereby realizing high-precision prediction of the deformation amount caused by rock mass structural rupture.

[0019]

[0020] ​In this embodiment, optionally, feature extraction is performed on the strain signal, including: The strain signal is subjected to differential signal processing, and features are extracted from the processed strain signal.

[0021] Specifically, a dynamic pitch adjustment technique can be used to spirally deploy helical wound optical fibers along the tunnel axis. This involves dynamically adjusting the pitch of the fiber across the entire strain field based on lithology and structure. The pitch values, arranged from largest to smallest, should be: stable rock strata pitch > mudstone section pitch > fault zone densified pitch, ultimately ensuring that the entire strain field covers the optical fiber. An epoxy resin-carbon nanotube composite coating is applied to the outside of the optical fiber as a coupling agent, and a micro-hydraulic tensioning clamp is used to achieve dynamic pre-tension control, ensuring tight coupling between the optical fiber and the surrounding rock.

[0022] The strain signals of the tunnel rock mass are acquired in real time by deploying wound optical fibers. Differential signal processing can be performed on the acquired strain signals to eliminate strain drift caused by temperature fluctuations, thereby further improving prediction accuracy. The specific calculation formula for differential signal processing is as follows: ; in, The original strain signal, The strain signal is measured by a temperature-compensated fiber optic cable. This is the strain signal after differential processing. After differential processing the acquired original strain signal, feature extraction can be performed on the differential strain signal.

[0023] In this embodiment, optionally, feature extraction is performed on the strain signal, including: A convolutional neural network is used to extract features from the strain signal.

[0024] Specifically, convolutional neural networks can be used to extract features from the differentially processed strain signal to filter out interference from construction machinery vibration and further improve prediction accuracy. Specifically, a three-layer convolutional neural network can be designed to extract the spatial features of the strain field from the strain signal. The specific calculation formula for the convolution kernel of the convolutional neural network is as follows: ; in, The time series data of the strain signal after differential processing represents the time series data of the strain signal. The time step, the first The strain value at each spatial location. These are the weight coefficients of the convolution kernel. This represents the size of the convolution kernel, i.e., the time step. The offset is set to adjust the baseline for output activation. This represents the activation function, used to introduce nonlinearity, preserve effective features, and suppress negative noise. The feature map extracted by the convolutional neural network is reduced in dimension by a GAP layer to obtain a strain time sequence feature.

[0025] In this embodiment, the displacement data is optionally subjected to feature extraction, including: The displacement data is subjected to notch filtering processing, and the processed displacement data is subjected to feature extraction.

[0026] Specifically, micro inertial measurement unit nodes can be implanted in the surrounding rock at a fixed interval to form a high-density monitoring grid, and the displacement data at different nodes of the tunnel rock mass is monitored in real time through the micro inertial measurement unit nodes. After obtaining the displacement data at the nodes, the existing notch filtering processing technology can be used to filter the displacement data, generate extremely low gain or phase inversion at the notch frequency corresponding to the mechanical vibration generated by the tunnel construction, and make the signal corresponding to the mechanical vibration frequency attenuate greatly, so as to suppress the mechanical interference caused by the tunnel construction and further improve the prediction accuracy.

[0027] In this embodiment, the displacement data is optionally subjected to feature extraction, including: The node displacement data of all nodes in the displacement data is fused by using a federated Kalman filtering algorithm.

[0028] Specifically, after the displacement data of the nodes is filtered, the federated Kalman filtering algorithm can be used to fuse the processed displacement data of all nodes to obtain the three-dimensional deformation characteristics of the tunnel rock mass.

[0029] Specifically, the state equation of the federated Kalman filtering algorithm is: ; Wherein, is the state vector, including the displacement, velocity and acceleration of the node. is the state transition matrix, is the control matrix, is the process noise, is the fused displacement, is the time.

[0030] The observation equation is: ; Wherein, is the observation value of the node, is the observation matrix, is the observation noise.

[0031] The federated Kalman filtering algorithm decomposes the global task into multiple local filters according to the information distribution principle, and then performs global fusion. The core steps include local filtering and global filtering. First, the global information is distributed to the local filters according to a certain rule Local filters: ; in, For the first Information allocation coefficients for each local filter.

[0032] Then, each local filter uses the assigned information and its own observation data to perform standard Kalman filtering independently and in parallel, obtaining the local error covariance matrix corresponding to the local state estimate.

[0033] Finally, the global filter collects the results of all local filters and fuses them according to the principle of "information conservation" to obtain the optimal global state estimate and global error covariance matrix. The specific calculation formulas for global fusion are as follows: ; ; The global error covariance matrix, This is for global state estimation. For the first The local error covariance matrix obtained from each local filter can automatically adjust the node weights. This is for local state estimation.

[0034] In the federated Kalman filter algorithm, local filtering operates independently, and single-node failures do not affect global fusion, thus exhibiting higher fault tolerance. Furthermore, local and global filtering are processed in parallel, improving computational efficiency. The algorithm can also automatically allocate node weights through the local error covariance matrix, assigning higher confidence to contour deformation zones along fault lines.

[0035] In this embodiment, optionally, feature extraction is performed on the AE signal, including: Wavelet transform is used to extract features from the AE signal.

[0036] Specifically, multiple acoustic emission (AE) sensors with a resonant frequency of 150kHz can be deployed per square meter in the tunnel to form an acoustic emission sensor array. This array can acquire AE signals within the tunnel in real time, and Morlet wavelet transform can be used to extract wavelet coefficients in the 50-100kHz characteristic frequency band as the time-series characteristics of the AE signals. The specific calculation formula is as follows: ; ; in, These are wavelet coefficients. , respectively, the scale parameter is used to control the stretching of the wavelet, which is inversely proportional to the frequency. The translation parameter is used to control the position of the wavelet on the time axis. is a complex exponential kernel, which is used to provide phase information to support the arrival-time difference positioning of the burst signal, is a complex exponential kernel, which is used to provide phase information to support the arrival-time difference positioning of the burst signal, is a Gaussian envelope, which is used for optimal time-frequency focusing. is the number of sample points, is the center frequency of the wavelet function, which determines the main energy distribution of the wavelet in the frequency domain.

[0037] The time-frequency localization characteristics of the Gaussian window in the Morlet wavelet transform can simultaneously capture both the burst-type burst signals such as rockburst microseisms and the continuous noise such as mechanical vibrations. It has higher time resolution than the short-time Fourier transform. And it can extract the phase difference in the AE signal to locate the position of the burst source. By adjusting the scale parameter to match the characteristic frequency band of soft rock rupture, it can effectively suppress low-frequency construction noise and further improve the prediction accuracy.

[0038] Through multiple tests, it is found that compared with the Daubechies wavelet transform and the short-time Fourier transform algorithm, the Morlet wavelet transform adopted in the embodiment has higher burst recognition rate and noise suppression effect, and lower positioning error. The specific verification results are as follows: In the embodiment, a 3D laser scanner can be used to periodically scan the tunnel rock mass to obtain deformation data of the tunnel rock mass.

[0039] After the multi-source time sequence feature vector is extracted, the LSTM prediction model trained can be used to distinguish the rock mass structure rupture event and predict the deformation amount caused by the rock mass structure rupture. The LSTM prediction model controls the information flow through the gating mechanism, including the forget gate, the input gate, the cell state update and the output gate. Among them, the forget gate determines the information lost in the cell state at the last time, and the specific calculation formula is as follows: ; wherein, is a multi-source time sequence feature vector, which contains the time sequence features of strain, displacement, AE signal and deformation data, is the hidden state at the last time, is the forget gate weight matrix, is the forget gate bias term, is a Sigmoid function.

[0040] The input gate determines the information in the current input that needs to be updated to the cell state, and the specific calculation formula is as follows: ; ; wherein, is the output of the input gate, is the candidate cell state, , are weight matrices corresponding to the input gate and the candidate state respectively, , are corresponding bias terms respectively, is a hyperbolic tangent activation function.

[0041] The long-term memory is updated in combination with the forget gate and the input gate, and the specific calculation formula is as follows: The output gate determines the information output to the hidden state in the cell state, and the specific expression is as follows: ; ; ; wherein, denotes a Hadamard product, , are the cell states output by the forget gate and the input gate respectively, is the candidate cell state, is the predicted cell state. is the predicted deformation of the tunnel rock mass, and the specific calculation formula is as follows: ; is an output layer weight matrix, is a final hidden state, is an output layer bias term.

[0042] In this embodiment, optionally, the deformation of the tunnel soft rock is predicted through the LSTM prediction model, which comprises: adjusting the contribution weight of each time sequence feature in the multi-source time sequence feature vector according to the lithology of the tunnel soft rock.

[0043] Specifically, adjusting the contribution weight can make the prediction model adapt to the local (rock) conditions and dynamically focus on the most sensitive and effective monitoring signals under the current geological conditions. The contribution weight is used inside the LSTM model, and by differentiating the weighting of the multi-source input features, it directly affects the importance of different data in the fusion analysis and final prediction of the model, thereby realizing more accurate and reliable deformation prediction.

[0044] In this embodiment, optionally, it further comprises: matching the predicted deformation with a multi-level early warning triggering condition to trigger a corresponding early warning signal.

[0045] Specifically, the deformation amount includes rock mass displacement prediction data and AE energy peak prediction value. After obtaining the deformation amount of the tunnel rock mass through the LSTM prediction model, the deformation amount can be compared with the trigger conditions corresponding to different warning levels, and the corresponding warning signal can be triggered according to the comparison result. For example, when the displacement prediction data of the tunnel rock mass exceeds the displacement threshold value 5mm / d, or the AE energy peak exceeds the peak threshold value , a yellow warning is triggered. When the displacement prediction data exceeds the displacement threshold value 10mm / d, and the AE energy peak exceeds the peak threshold value , an orange warning is triggered. When the displacement prediction data exceeds the displacement threshold value 15mm / d, and the AE energy peak mutation rate exceeds , a red warning is triggered.

[0046] A real-time monitoring system for tunnel soft rock deformation, comprising: a data acquisition module configured to acquire strain signals, displacement data, AE signals and deformation data of the tunnel soft rock; a feature extraction module configured to extract features from the strain signals, displacement data, AE signals and deformation data respectively, and obtain corresponding time sequence features to form a multi-source time sequence feature vector; a deformation prediction module configured to predict the deformation amount of the tunnel soft rock caused by rock mass structural rupture through an LSTM prediction model according to the multi-source time sequence feature vector.

[0047] Specifically, the monitoring system includes a data acquisition module, a feature extraction module and a deformation prediction module. The data acquisition module can acquire strain signals, displacement data, AE signals and deformation data of the tunnel soft rock through corresponding monitoring devices arranged in the tunnel. The feature extraction module can extract strain time sequence features, displacement time sequence features, AE signal time sequence features and deformation time sequence features from the strain signals, displacement data, AE signals and deformation data respectively. The strain time sequence features, displacement time sequence features, AE signal time sequence features and deformation time sequence features are aligned and combined into a multi-source time sequence feature vector. The deformation prediction module can input the multi-source time sequence feature vector into the trained LSTM prediction model, and fuse and analyze the four-dimensional feature vector in the multi-source time sequence feature vector through the LSTM prediction model, so as to accurately distinguish between rock mass deformation caused by structural rupture and rock mass deformation caused by mechanical vibration, and accurately predict the displacement amount of rock mass deformation, thereby realizing high-precision prediction of the deformation amount caused by rock mass structural rupture.

[0048] The above examples are only used to illustrate the technical solutions of the present application, but not limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some or all of the technical features can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the claims and the specification of the present application.

Claims

1. A method for real-time monitoring of deformation of soft rock in a tunnel, characterized in that, The method comprises the following steps: obtaining strain signals, displacement data, AE signals and deformation data of the tunnel soft rock; performing feature extraction on the strain signals, displacement data, AE signals and deformation data respectively to obtain corresponding time sequence features to form a multi-source time sequence feature vector; predicting the deformation amount of the tunnel soft rock caused by structural rupture of the rock mass by an LSTM prediction model according to the multi-source time sequence feature vector.

2. The method according to claim 1, wherein, The feature extraction on the strain signals comprises the following steps: performing differential signal processing on the strain signals and performing feature extraction on the processed strain signals.

3. The method of claim 1, wherein, The feature extraction on the strain signals comprises the following steps: performing feature extraction on the strain signals by using a convolutional neural network.

4. The method of claim 1, wherein, The feature extraction on the displacement data comprises the following steps: performing notch filter processing on the displacement data and performing feature extraction on the processed displacement data.

5. The method of claim 1, wherein, The feature extraction on the displacement data comprises the following steps: fusing node displacement data of all nodes in the displacement data by using a federated Kalman filter algorithm.

6. The method of claim 1, wherein, The feature extraction on the AE signals comprises the following steps: performing feature extraction on the AE signals by using a wavelet transform.

7. The method of claim 1, wherein, The prediction of the deformation amount of the tunnel soft rock by the LSTM prediction model comprises the following steps: adjusting the contribution weight of each time sequence feature in the multi-source time sequence feature vector according to the lithology of the tunnel soft rock.

8. The method for real-time monitoring of deformation of soft rock in a tunnel according to claim 1, characterized in that, The method further comprises the following steps: matching the predicted deformation amount with multi-level early warning triggering conditions to trigger corresponding early warning signals.

9. A real-time monitoring system for deformation of soft rock in a tunnel, characterized in that, The method comprises the following steps: a data acquisition module configured to obtain strain signals, displacement data, AE signals and deformation data of the tunnel soft rock; a feature extraction module configured to perform feature extraction on the strain signals, displacement data, AE signals and deformation data respectively to obtain corresponding time sequence features to form a multi-source time sequence feature vector; a deformation prediction module configured to predict the deformation amount of the tunnel soft rock caused by structural rupture of the rock mass by an LSTM prediction model according to the multi-source time sequence feature vector.

Images