Deep learning-based tunnel face surveying method

Abstract

The invention discloses a tunnel face surveying method based on deep learning. The method comprises the following steps: acquiring high-resolution point cloud data of a tunnel face and a peripheral area of the tunnel face, and preprocessing the data; automatically segmenting a tunnel face area and a surrounding area, and extracting tiny deformation features from segmented face point clouds; constructing a fusion prediction model FDNN of a CNN convolutional neural network and an LSTM long short-term memory network, extracting spatial features of the tunnel face point cloud, processing time dependence of point cloud deformation data, and capturing dynamic changes of the tunnel face during construction; carrying out model training to optimize a loss function, adjusting model parameters and carrying out multi-round iterative training; and inputting the real-time point cloud data into the trained fusion prediction model FDNN, performing tunnel face deformation trend prediction, and presetting a deformation safety threshold. According to the scheme, comprehensive learning of deformation features is realized, fine temporal and spatial changes of the tunnel face are captured, and the monitoring effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel construction monitoring, and in particular to a tunnel face survey method based on deep learning. Background Art

[0002] Existing technologies can be divided into two main categories: traditional monitoring methods based on physical sensors and preliminary intelligent monitoring methods based on machine learning.

[0003] (1) Traditional methods based on physical sensors: Currently, most tunnel monitoring methods use a variety of physical sensors, such as laser displacement sensors, strain gauges, and pressure sensors, to obtain data such as tunnel deformation, displacement, and pressure. These sensors are deployed on the tunnel face and surrounding support structures, and the degree of tunnel face deformation is determined by regularly collecting data. Data processing relies on manual or simple statistical analysis to perform trend analysis on monitoring data.

[0004] (2) Preliminary intelligent monitoring methods based on machine learning: With the development of data-driven technology, some monitoring systems have begun to introduce machine learning methods, such as support vector machines (SVM) and random forests. These methods are mainly used for offline analysis of deformation data and predict deformation trends through regression analysis of historical data. Such models can automate data analysis to a certain extent and improve data utilization.

[0005] However, the methods in the prior art have the following defects:

[0006] (1) Traditional methods have low monitoring frequency and data processing efficiency under complex geological conditions, and often cannot meet the high demands for real-time and accuracy during construction.

[0007] (2) The data may contain errors or lags: Due to the high variability of the construction environment and the limited distribution of sensors, the data may contain errors or lags, making it difficult to provide stable monitoring results.

[0008] (3) Limited ability to process time series data: and it is difficult to cope with the complex nonlinear changes in face data. Therefore, when faced with sudden geological changes, machine learning methods are still insufficient in terms of real-time and adaptability, and it is difficult to cope with the needs of face deformation monitoring in complex environments in real time.

[0009] (4) Manual threshold setting is required: Most existing methods require manual threshold setting to determine whether there is an anomaly. Threshold setting is difficult and requires repeated trials and adjustments. Summary of the invention

[0010] In response to the problems of low monitoring accuracy, inability to automatically identify abnormal trends, poor real-time performance, inflexible thresholds, etc. in tunnel construction face deformation survey, the present invention provides a face survey method based on deep learning, which introduces deep learning technology into the automatic monitoring and trend prediction framework of face deformation, fully combines the spatial information and time information of point cloud data, establishes a deformation trend model, identifies anomalies through deformation prediction errors, and solves the problems mentioned in the above background technology.

[0011] The present invention is implemented by adopting the following technical scheme: a tunnel face survey method based on deep learning, comprising the following steps:

[0012] Step S1: Scan the tunnel face, obtain high-resolution point cloud data of the tunnel face and its surrounding areas, and pre-process the data;

[0013] Step S2: Automatically segment the tunnel face area and the surrounding area, and extract tiny deformation features from the segmented tunnel face point cloud;

[0014] Step S3: construct a fusion prediction model FDNN of CNN convolutional neural network and LSTM long short-term memory network, use CNN convolutional neural network to extract spatial features of tunnel face point cloud, use LSTM long short-term memory network to process the time dependency of point cloud deformation data, and capture the dynamic changes of tunnel face during construction;

[0015] Step S4: Use historical tunnel face data to perform model training, optimize the loss function, adjust model parameters, and perform multiple rounds of iterative training;

[0016] Step S5: input the real-time point cloud data into the trained fusion prediction model FDNN to predict the deformation trend of the tunnel face and preset the deformation safety threshold.

[0017] Specifically, the tunnel face scanning uses a three-dimensional laser scanner, and the laser scanning frequency and sampling density are set according to the geometric characteristics of the tunnel face and the monitoring accuracy requirements.

[0018] Specifically, the data preprocessing step S1 includes the following sub-steps:

[0019] Step S11: using a filtering algorithm to remove point cloud noise;

[0020] Step S12: Eliminate outliers that deviate from a large range according to statistical methods;

[0021] Step S13: introduce a stable rigid reference object and use a multi-level iterative registration algorithm to align the point clouds of each period; the formula for point cloud registration is:

[0022] T = RP + t;

[0023] Among them, T is the target point after registration, R is the rotation matrix, P is the initial point cloud, and t is the translation vector; iteratively optimize R and t to minimize the mean square error of all point clouds. The formula is:

[0024] min R,t ∑ i |RP i +tQ i | 2 ;

[0025] Among them, P i and Q i are the corresponding points of the source point cloud and the target point cloud, R and t are the rotation matrix and translation vector respectively.

[0026] Specifically, the face region segmentation in step S2 uses principal component analysis to calculate the normal vector, and extracts the face point cloud based on the angle between the normal vector direction and the tunnel axis direction. The normal vector is calculated by calculating the covariance matrix C of the point cloud data and obtaining the eigenvector ω corresponding to its minimum eigenvalue. The specific calculation formula is:

[0027]

[0028] Here, cov(x,y) represents the covariance of the x and y coordinates.

[0029] Specifically, the deformation feature is extracted using two-dimensional continuous wavelet transform, and the calculation formula is:

[0030]

[0031] Among them, ψ a,b,s (x, y) is the mother wavelet function, a and b are the translation parameters, s is the scale factor, and g(x, y) is the deformation signal of the tunnel face.

[0032] Specifically, the step S3 uses a CNN convolutional neural network to extract spatial features, and the expression is:

[0033] (f*g)(i,j)=∑ m ∑ n f(m,n)·g(im,jn);

[0034] Among them, f is the convolution kernel and g is the input feature image.

[0035] Specifically, the LSTM long short-term memory network takes the features extracted by CNN as input, identifies long-term and short-term trends in the deformation sequence, and predicts the deformation trend in the form of time series, including the following parts:

[0036] Forget Gate t :Control the memory cell state of the previous time step (Ct-1 ) The number of times that need to be retained until the current time step, the forget gate f t The output is in [0,1], 1 means completely retained, 0 means completely discarded; the calculation formula is:

[0037] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0038] Among them, W f is the weight matrix of the forget gate, b f is the bias vector, σ is the sigmoid activation function;

[0039] Input Gate i t : Controls the input x of the current time step t and the generated candidate states The part that should be added to the memory unit is output as [0,1], and the calculation formula is:

[0040] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0041] Among them, W i is the weight matrix of the input gate, b i is the bias vector;

[0042] Candidate memory cell status Generated by the current input and the previous hidden state, it serves as the main source of new information for the current time step of the LSTM long short-term memory network; candidate memory unit state Generated by the tanh activation function, the calculation formula is:

[0043]

[0044] Among them, W C is the weight matrix of the candidate state, b C is the bias vector.

[0045] Specifically, the LSTM long short-term memory network also includes:

[0046] Memory cell status C t : Store long-term information through the forget gate f t and input gate i t The information u is retained and updated by regulation, and the update formula is:

[0047]

[0048] Among them, f t ×C T-1 Indicates the last state that is retained. Indicates new information added to the current time step;

[0049] Output gate o t :Determine from the current memory cell state C t Output to hidden state h t The information is used for the calculation of the next time step. The calculation formula is:

[0050] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0051] Among them, W o is the weight matrix of the output gate, b o is the bias vector, x t is the input data of the current time step t, h t-1 is the hidden state of the previous time step t-1;

[0052] Hidden state h t : As the final output, it contains the deformation prediction information of the current time step, which will be passed to the next time step for time series prediction; the current memory unit state C t and output gate o t Determine, the calculation formula is:

[0053] h t =o t ×tanh(C t ).

[0054] Specifically, the optimization loss function formula in step S4 is:

[0055]

[0056] Among them, y i is the actual deformation data, is the predicted value, and N is the number of samples.

[0057] The beneficial effects of the present invention are as follows: the present invention uses a deep learning model to perform intelligent monitoring and early warning analysis of tunnel face deformation to ensure the safety of the structure during underground construction. Through data-driven automatic feature extraction, the deep learning model can automatically extract high-dimensional nonlinear features from the data, reduce human errors, and improve monitoring accuracy; through the early warning mechanism based on the fusion deep neural network (FDNN), it can automatically trigger an alarm by setting a threshold (such as displacement exceeding the standard value) to ensure construction safety; by continuously inputting new construction monitoring data, the model can be trained online and parameters can be continuously updated to adapt to different geological conditions and construction conditions. Traditional monitoring methods mainly rely on physical sensors and periodic manual measurements, and it is difficult to respond to rapidly changing construction environments in real time, especially under complex geological conditions. Therefore, by introducing deep learning technology, a large amount of sensor data can be analyzed more efficiently to achieve real-time monitoring and automatic early warning of tunnel face deformation. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0059] Figure 1 This is a flow chart of a tunnel face survey method based on deep learning in an embodiment of the present invention;

[0060] Figure 2 This is a diagram of the LSTM long short-term memory network architecture in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0062] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0063] The following is combined with Figure 1-2 , some embodiments of the present invention are described in detail. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0064] In order to solve the problems of low monitoring accuracy, inability to automatically identify abnormal trends, weak real-time performance, and inflexible thresholds in tunnel construction face deformation survey, the present invention provides a face deformation survey method based on deep learning, introduces deep learning technology into the automatic monitoring and trend prediction framework of face deformation, fully combines the spatial information and time information of point cloud data, establishes a deformation trend model, and identifies anomalies through deformation prediction errors, thus solving the problems mentioned in the above background technology. The method realizes comprehensive data collection of the face through three-dimensional laser scanning technology; constructs spatial feature extraction and time-dependent modeling of the face area through a deep learning network; monitors abnormal changes of the face at different time steps through fusion prediction error analysis; realizes noise-resistant deformation anomaly detection through the setting of flexible thresholds; and locates abnormal moments and positions through errors to provide explanatory analysis for deformation trends of the face. In a preferred embodiment, a face survey method based on deep learning includes the following steps:

[0065] Step S1: Scan the tunnel face, obtain high-resolution point cloud data of the tunnel face and its surrounding areas, and pre-process the data;

[0066] Step S2: Automatically segment the tunnel face area and the surrounding area, and extract tiny deformation features from the segmented tunnel face point cloud;

[0067] Step S3: construct a fusion prediction model FDNN of CNN convolutional neural network and LSTM long short-term memory network, use CNN convolutional neural network to extract spatial features of tunnel face point cloud, use LSTM long short-term memory network to process the time dependency of point cloud deformation data, and capture the dynamic changes of tunnel face during construction;

[0068] Step S4: Use historical tunnel face data to perform model training, optimize the loss function, adjust model parameters, and perform multiple rounds of iterative training;

[0069] Step S5: input the real-time point cloud data into the trained fusion prediction model FDNN to predict the deformation trend of the tunnel face and preset the deformation safety threshold.

[0070] In another embodiment, a tunnel face deformation survey method based on deep learning, such as Figure 1 The detailed steps are as follows:

[0071] Step 1: Data collection and preprocessing, including:

[0072] Step 1.1 Data acquisition: Use a 3D laser scanner to scan the tunnel face and obtain high-resolution point cloud data of the tunnel face and its surrounding areas. The laser scanning frequency and sampling density are set according to the geometric characteristics of the tunnel face and the monitoring accuracy requirements to ensure that the data coverage is comprehensive and accurate.

[0073] Step 1.2 Data preprocessing: Process the collected raw point cloud data, including denoising, outlier cleaning, point cloud registration and other steps to ensure the quality and consistency of the monitoring data. In this embodiment:

[0074] (1) Denoising: Use median filtering or bilateral filtering algorithm to remove point cloud noise.

[0075] (2) Outlier cleaning: Based on statistical methods (such as box plots), outliers that deviate from a large range are eliminated.

[0076] (3) Point cloud registration: introduce a stable rigid reference object (such as initial support) and use a multi-level iterative registration algorithm to align the point clouds of each period. The core formula of point cloud registration is as follows:

[0077] T=RP+t

[0078] Among them, T is the target point after registration, R is the rotation matrix, P is the initial point cloud, and t is the translation vector.

[0079] Iteratively optimize R and t to minimize the mean square error of all point clouds:

[0080]

[0081] Among them, P i and Q i are the corresponding points of the source point cloud and the target point cloud, R and t are the rotation matrix and translation vector respectively.

[0082] Step 2: Segmentation and feature extraction of the tunnel face, including:

[0083] Step 2.1 Segmentation of tunnel face area: The tunnel face area and the surrounding area are automatically segmented by normal vector clustering. The normal vector is calculated using the principal component analysis (PCA) method, and the tunnel face point cloud is extracted based on the angle between the normal vector direction and the tunnel axis direction.

[0084] In this embodiment, PCA normal vector calculation: calculate the covariance matrix (C) of the point cloud data and obtain the eigenvector (ω) corresponding to its minimum eigenvalue, which is the normal vector:

[0085]

[0086] Where (cov(x,y)) represents the covariance of the (x) and (y) coordinates.

[0087] Step 2.2 Extraction of features of the tunnel face point cloud data: Use two-dimensional continuous wavelet transform (CWT) to extract tiny deformation features from the segmented tunnel face point cloud.

[0088] In this embodiment, the two-dimensional wavelet transform formula is:

[0089]

[0090] Among them, (ψ a,b,s (x,y)) is the mother wavelet function, (a,b) is the translation parameter, (s) is the scale factor, and (g(x,y)) is the deformation signal of the tunnel face. The local undulation characteristics of each sub-area of ​​the tunnel face are obtained through wavelet transform for subsequent deformation trend analysis.

[0091] Step 3: Deformation analysis model based on deep learning: Construct a fusion model (FDNN) of convolutional neural network (CNN) and long short-term memory network (LSTM) to process the spatial characteristics and time series characteristics of the face.

[0092] (1) CNN convolution layer: Use convolutional neural network to extract the spatial features of the tunnel face point cloud. The convolution operation is used to analyze the features of different regions in the tunnel face data. The formula is as follows:

[0093]

[0094] Among them, (f) is the convolution kernel and (g) is the input feature image.

[0095] (2) LSTM time series layer: Use the long short-term memory network (LSTM) to process the time dependency of point cloud deformation data and capture the dynamic changes of the tunnel face during construction; Figure 2 As shown, the features extracted by CNN are used as input to predict the deformation trend in the form of time series. In this embodiment, the gate control unit formula of LSTM is as follows:

[0096] f t :Forget Gate. The forget gate controls the state of the memory unit at the previous time step (C t-1 )) How many need to be retained to the current time step. The output of the forget gate is between [0,1], 1 means completely retained, and 0 means completely discarded. The calculation formula of the forget gate is:

[0097] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0098] Among them, W f is the weight matrix of the forget gate, b f is the bias vector and σ is the sigmoid activation function.

[0099] it : Input Gate. The input gate controls the input x of the current time step. t and the generated candidate states Which parts of the input gate should be added to the memory unit. The output of the input gate is also between [0,1]. The calculation formula of the input gate is:

[0100] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0101] Among them, W i is the weight matrix of the input gate, b i is the bias vector.

[0102] : Candidate cell state. The candidate state is generated by the current input and the previous hidden state, and is the main source of new information for the current time step of the LSTM unit. The candidate state is usually generated by the tanh activation function, and its calculation formula is:

[0103]

[0104] Among them, W C is the weight matrix of the candidate state, b C is the bias vector.

[0105] C t :Cell State. The cell state stores long-term information in LSTM, and the information is retained and updated through the regulation of the forget gate and input gate. The update formula is:

[0106]

[0107] Among them, f t ×C t-1 Indicates the last state that is retained. Indicates the new information added to the current time step.

[0108] o t :Output Gate. The output gate determines which information is output from the current memory cell state C t Output to hidden state h t , used for calculation of the next time step. The calculation formula of the output gate is:

[0109] o t =σ(W o ·[h t-1 ,xt ]+b o )

[0110] Among them, W o is the weight matrix of the output gate, b o is the bias vector.

[0111] h t :Hidden State. The hidden state is the final output of LSTM, which contains the deformation prediction information of the current time step. The hidden state is composed of the current memory cell state C t and output gate o t Decide:

[0112] h t =o t ×tanh(C t )

[0113] The hidden state is passed to the next time step for time series prediction.

[0114] x t : Input data at the current time step t. Usually a vector, representing the monitoring data of the tunnel face or other systems at the current moment.

[0115] h t-1 : The hidden state of the previous time step t-1. The hidden state is one of the outputs of LSTM, which contains the context of the time series and is used to capture the characteristics of data over time.

[0116] The LSTM layer uses these formulas to identify long-term and short-term trends in the deformation sequence, providing information for tunnel face deformation prediction.

[0117] Step 4: Model training and optimization: Use historical face data to perform model training. In this embodiment, the loss function (such as mean square error MSE) is optimized:

[0118]

[0119] Among them, (y i ) is the actual deformation data, is the predicted value, (N) is the number of samples.

[0120] The Adam optimizer is used to adjust the model parameters, and the accuracy of model prediction is improved through multiple rounds of training iterations.

[0121] Step 5: Real-time monitoring and intelligent early warning, including:

[0122] Real-time data input: input the real-time point cloud data into the trained FDNN model to predict the deformation trend of the tunnel face.

[0123] Early warning mechanism: The system presets a deformation safety threshold (such as maximum displacement). When the deformation value predicted by the model exceeds the safety threshold, an early warning signal is triggered to notify the construction personnel.

[0124] The key points of the present invention include:

[0125] 1. High-precision point cloud acquisition and processing: Full coverage of the tunnel face and high-precision data acquisition are achieved through 3D laser scanning.

[0126] 2. Automatic feature extraction of deep learning models: Using convolutional neural networks and long short-term memory networks, the spatial and temporal features of deformation are automatically extracted from point cloud data.

[0127] 3. Intelligent real-time warning: Through deep learning, deformation trends are predicted and warnings are automatically triggered when abnormal deformation is detected to ensure construction safety.

[0128] The present invention discloses a tunnel face deformation survey method based on deep learning, which belongs to the field of intelligent monitoring of tunnel engineering. It aims to realize automatic identification and prediction of tunnel face deformation through deep learning, timely discover abnormal deformation that may occur during tunnel construction, and ensure engineering safety.

[0129] The spatial and temporal characteristics of the face point cloud data are used to improve the accuracy and real-time performance of deformation monitoring. Through the spatial information extraction of 3D laser scanning data and the time series prediction of the deep learning model, this method can identify the slight deformation trend of the face and provide refined monitoring for deformation detection of the face. In view of the problems of limited monitoring range and slow response in traditional methods, the present invention uses a deep learning model combining CNN and LSTM to achieve comprehensive learning of deformation characteristics, capture subtle spatiotemporal changes of the face, and improve the monitoring effect.

[0130] This solution also sets up an intelligent early warning mechanism, which greatly improves the accuracy and noise resistance of anomaly detection. Through flexible threshold setting and aggregate error analysis, various types of anomaly detection of face deformation are achieved, reducing misjudgments caused by noise interference. Compared with the traditional method of manually setting thresholds, the adaptive threshold method of the present invention can effectively identify real abnormal deformations in noisy environments, ensuring the safe response of construction personnel.

[0131] The present invention has high computational efficiency and excellent interpretability. By analyzing the prediction error and locating the abnormal area, the system can provide detailed abnormal explanations to help construction personnel determine the time and specific location of the abnormality and take preventive measures in a timely manner. Compared with traditional face monitoring methods, the present invention realizes the automation of monitoring and analysis through a deep learning model, greatly improves the monitoring efficiency, and has good prediction accuracy and fast abnormal response speed.

[0132] For the aforementioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily required by the present application.

[0133] The above embodiments describe the basic principles and main features of the present invention and the advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the changes and modifications made by those skilled in the art shall be within the scope of protection of the appended claims of the present invention without departing from the spirit and scope of the present invention.

Claims

1. A tunnel face survey method based on deep learning, characterized in that: The following steps are involved: Step S1: Scan the tunnel face, obtain high-resolution point cloud data of the tunnel face and its surrounding areas, and pre-process the data; Step S2: Automatically segment the tunnel face area and the surrounding area, and extract tiny deformation features from the segmented tunnel face point cloud; Step S3: construct a fusion prediction model FDNN of CNN convolutional neural network and LSTM long short-term memory network, use CNN convolutional neural network to extract spatial features of tunnel face point cloud, use LSTM long short-term memory network to process the time dependency of point cloud deformation data, and capture the dynamic changes of tunnel face during construction; Step S4: Use historical tunnel face data to perform model training, optimize the loss function, adjust model parameters, and perform multiple rounds of iterative training; Step S5: input the real-time point cloud data into the trained fusion prediction model FDNN to predict the deformation trend of the tunnel face and preset the deformation safety threshold.

2. A tunnel face survey method based on deep learning as claimed in claim 1, characterized in that: The tunnel face scanning uses a three-dimensional laser scanner, and the laser scanning frequency and sampling density are set according to the geometric characteristics of the tunnel face and the monitoring accuracy requirements.

3. A tunnel face survey method based on deep learning as claimed in claim 2, characterized in that: The step S1 data preprocessing includes the following sub-steps: Step S11: using a filtering algorithm to remove point cloud noise; Step S12: Eliminate outliers that deviate from a large range according to statistical methods; Step S13: introduce a stable rigid reference object and use a multi-level iterative registration algorithm to align the point clouds of each period; the formula for point cloud registration is: T = RP + t; Among them, T is the target point after registration, R is the rotation matrix, P is the initial point cloud, and t is the translation vector; iteratively optimize R and t to minimize the mean square error of all point clouds. The formula is: min R,t ∑ i |RP i +t-Q i | 2 ; Among them, P i and Q i are the corresponding points of the source point cloud and the target point cloud, R and t are the rotation matrix and translation vector respectively.

4. The method for tunnel face survey based on deep learning according to claim 1, characterized in that: In step S2, the normal vector is calculated by principal component analysis method for the tunnel face region segmentation, and the tunnel face point cloud is extracted based on the angle between the normal vector direction and the tunnel axis direction. The normal vector is calculated by calculating the covariance matrix C of the point cloud data and obtaining the eigenvector ω corresponding to its minimum eigenvalue. The specific calculation formula is: Here, cov(x,y) represents the covariance of the x and y coordinates.

5. The method for tunnel face survey based on deep learning according to claim 4, characterized in that: The deformation feature is extracted using two-dimensional continuous wavelet transform, and the calculation formula is: Among them, ψ a,b,s (x, y) is the mother wavelet function, a and b are the translation parameters, s is the scale factor, and g(x, y) is the deformation signal of the tunnel face.

6. The method for tunnel face survey based on deep learning according to claim 1, characterized in that: The step S3 uses the CNN convolutional neural network to extract spatial features, and the expression is: (f*g)(i,j)=∑ m ∑ n f(m,n)·g(i-m,j-n); Among them, f is the convolution kernel and g is the input feature image.

7. The method for tunnel face survey based on deep learning according to claim 1, characterized in that: The LSTM long short-term memory network takes the features extracted by CNN as input, identifies long-term and short-term trends in the deformation sequence, and predicts the deformation trend in the form of time series, including the following parts: Forget Gate t :Control the memory cell state of the previous time step (C t-1 ) The number of times that need to be retained until the current time step, the forget gate f t The output is in [0,1], 1 means completely retained, 0 means completely discarded; the calculation formula is: f t =σ(W f ·[h t-1 ,x t ]+b f ); Among them, W f is the weight matrix of the forget gate, b f is the bias vector, σ is the sigmoid activation function; Input Gate i t : Controls the input x of the current time step t and the generated candidate states The part that should be added to the memory unit is output as [0,1], and the calculation formula is: i t =σ(W i ·[h t-1 ,x t ]+b i ); Among them, W i is the weight matrix of the input gate, b i is the bias vector; Candidate memory cell status Generated by the current input and the previous hidden state, it serves as the main source of new information for the current time step of the LSTM long short-term memory network; candidate memory unit state Generated by the tanh activation function, the calculation formula is: Among them, W C is the weight matrix of the candidate state, b C is the bias vector.

8. The method for tunnel face survey based on deep learning according to claim 7, characterized in that: The LSTM long short-term memory network also includes: Memory cell state C t : Store long-term information through the forget gate f t and input gate i t The information u is retained and updated by regulation, and the update formula is: Among them, f t ×C t-1 Indicates the last state that is retained. Indicates new information added to the current time step; Output gate o t :Determine from the current memory cell state C t Output to hidden state h t The information is used for the calculation of the next time step. The calculation formula is: the t =σ(W o ·[h t-1 ,x t ]+b o ); Among them, W o is the weight matrix of the output gate, b o is the bias vector, x t is the input data of the current time step t, h t-1 is the hidden state of the previous time step t-1; Hidden state h t : As the final output, it contains the deformation prediction information of the current time step, which will be passed to the next time step for time series prediction; the current memory unit state C t and output gate o t Determine, the calculation formula is: h t =o t ×tanh(C t )。 9. The method for tunnel face survey based on deep learning according to claim 1, characterized in that: The optimization loss function formula in step S4 is: Among them, y i is the actual deformation data, is the predicted value, and N is the number of samples.

Images