Multi-factor tunnel surrounding rock deformation prediction method based on deep learning

Abstract

The invention discloses a multi-factor tunnel surrounding rock deformation prediction method based on deep learning, and the method effectively improves the accuracy of tunnel surrounding rock deformation prediction through integrating an information collaborative analysis index and a long-short-term memory network fused with an attention mechanism. According to the method, the LSTM model integrating the information theory and the attention mechanism is constructed, and the model can be dynamically focused on key time sequence data, so that the prediction accuracy is improved. In addition, the method also comprises a model verification step: comparing a prediction result with measured data, and evaluating the performance of the model by using a mean absolute error and a mean square error. The informatization and automation level of tunnel surrounding rock deformation prediction is improved, and powerful technical support is provided for tunnel construction safety.

Description

Background Art

[0001] With the rapid development of modern transportation networks, tunnel engineering, as one of the key infrastructures, plays a vital role in the complex and ever-changing geological environment. The geological structure in the Northwest region is particularly complex. During the excavation and subsequent support of tunnels, the stability of the surrounding rock is affected by many factors, such as the physical and mechanical properties of the rock mass, groundwater level changes, and ground stress distribution. These factors work together to cause the surrounding rock to deform or even be damaged, thus affecting the safety and service life of the tunnel structure.

[0002] Traditional tunnel surrounding rock deformation prediction methods mainly rely on empirical formulas and classical mechanical models, such as elasticity theory, plasticity theory, etc. However, these methods often fail to fully consider the various complex factors that affect the stability of the surrounding rock, and have limitations when processing large amounts of heterogeneous data. In addition, due to the uncertainty of the geological environment and the dynamic changes during the construction process, traditional methods have been unable to meet the needs of modern tunnel engineering in terms of prediction accuracy and real-time performance.

[0003] In recent years, with the rise of deep learning technology, its advantages in processing large-scale, high-dimensional data have gradually been applied to tunnel surrounding rock deformation prediction. The LSTM model has been widely used in deformation prediction due to its excellent performance in time series analysis. However, existing studies mostly focus on the single influencing factor of time series data, lacking comprehensive consideration of multivariate information in actual engineering, which limits the generalization ability and prediction accuracy of the model. Summary of the invention

[0004] In view of this, it is particularly urgent to develop a tunnel surrounding rock deformation prediction method that can integrate multivariate information and deep learning. The present invention aims to establish an LSTM model that integrates information collaborative analysis index (ISAI) and attention mechanism (MHA) through deep learning technology to achieve high-precision prediction of tunnel surrounding rock deformation. This method can not only improve the accuracy and reliability of prediction, but also provide a scientific basis for the safe construction and maintenance of tunnel engineering.

[0005] To achieve the above object, the present invention provides a tunnel surrounding rock deformation prediction method based on the ISAI-MHA-LSTM model, comprising the following steps:

[0006] S1. Multivariate information set data collection and processing;

[0007] S2. Construction of multivariate information set;

[0008] S3. Data pre-processing;

[0009] S4. Input the data set and divide the source data into a training set and a test set;

[0010] S5.ISAI-MHA-LSTM model construction;

[0011] S6. Define the LSTM layer training step size and adjust the model parameters;

[0012] S7. Select indicators to verify and evaluate the model;

[0013] Further, the step S1 includes the following sub-steps:

[0014] S1.1 Collect deformation data of tunnel surrounding rock, including peripheral convergence deformation data and arch crown settlement data of surrounding rock;

[0015] S1.2 Collect multivariate monitoring data of tunnel surrounding rock, including but not limited to geological conditions, environmental factors, etc.;

[0016] S1.3 Perform curve regression analysis on the collected multivariate information data and use the curve shape to reflect the monitoring status of the data;

[0017] S1.4 When the curve morphology shows abnormal volatility, immediately investigate the cause and take measures;

[0018] Further, the step S2 includes the following sub-steps:

[0019] S2.1 Based on actual engineering, seven information factors, namely water content, distance to section, cohesion, burial depth ratio, compression coefficient, friction angle and density, are selected to construct the initial multivariate information factor set (X′) = {x′ 1 , x′ 2 , ..., x′ 7};

[0020] S2.2 obtain the ISAI of each information factor;

[0021] The calculation method of ISAI is as follows:

[0022]

[0023] MIC Zi is the information factor x′ in the surrounding convergent deformation data i The calculated maximum mutual information coefficient, MIC Gi is the information factor x′ in the vault settlement data i The maximum mutual information coefficient calculated;

[0024] S2.3 Assign ISAI as a weight to each corresponding information factor, and finally form a multivariate factor information set (X) = {x 1 , x 2 ,......,x 8};

[0025] Furthermore, the step S3 includes the following sub-steps:

[0026] S3.1 Data cleaning, including removing duplicate records in the data and correcting or deleting obviously erroneous data entries;

[0027] S3.2 Outlier detection and processing, including identifying outliers in the data that do not conform to the norm. These outliers may be caused by measurement errors or other abnormal conditions. The processing methods include deletion, replacement or adjustment using statistical methods;

[0028] S3.3 Data standardization, that is, converting data to a unified scale to eliminate the impact of different dimensions and magnitudes, and normalizing the data to the range of [0,1];

[0029] Furthermore, the formula for data standardization is:

[0030]

[0031] Among them, a * is the value after data standardization, a is the original value of the data, min(G) is the minimum value of the original data set, and range(G) is the value range of the original data set;

[0032] Furthermore, the step S4 includes the following sub-steps:

[0033] S4.1 defines the input data set as D = {d 1 , d 2 ...d j ,...d n};

[0034] d j ={x 1j , x 2j , x 3j , x 4j , x 5j , x 6j , x 7j , x 8j}

[0035] Among them, d j represents the jth group of data;

[0036] S4.2 Divide the data after the influence factor correlation analysis and weight assignment into a training set D′ train and the test set D′ test ;

[0037] D′ train ={d 1 , d 2,…d j}

[0038] D′ test ={d j+1 , d j+2 , …d n}

[0039] According to the step size, 80% of the data set is used as the training set D′ train , 20% as the test set D′ test ; Further, the step S6 includes the following sub-steps:

[0040] S6.1 sets the hidden layer size and activation function, and adds the Dropout method to the LSTM layer;

[0041] S6.2 sets the model prediction interval to 1 and the input step size to 10, based on which the sliding prediction predict(x 10 : 10+1 )=output 10+1 ;

[0042] S6.3 In the test data set, 7 groups of samples are selected for final prediction, and these prediction results are saved;

[0043] Based on the above steps, the calculation process of the ISAI-MHA-LSTM model is as follows:

[0044] I. Set the data set D′={d 1 , d 2 , d t ,...d s},d t ={d′ t , d′ t+1 ...d′ j-s+t-1} Input the hidden layer of LSTM;

[0045] Among them, s is the window size for time series analysis, d′ t The data at time t

[0046] II. The output of the LSTM hidden layer is O′={o 1 , o 2 , o t ,... s},ot={o′ t , o′ t+1 ...o′ j-s+t-1};

[0047] III. Calculate the self-attention score in the MHA layer, output the weight matrix, and input the weighted data into the fully connected layer;

[0048] IV. The actual output after the fully connected layer is R = {r 1 , r 2 , r t ,...r s}

[0049] Furthermore, the step S7 includes the following sub-steps:

[0050] S7.1 Verify the prediction effect of the model and compare the predicted data with the measured data. The mean absolute error (MAE) and mean square error (MSE) are used to evaluate the model effect;

[0051] S7.2 obtains the real predicted data through denormalization and compares it with the measured data;

[0052] Furthermore, the calculation formulas for the mean absolute error (MAE) and mean square error (MSE) mentioned in S7 are as follows:

[0053]

[0054] in, is the actual deformation value of the surrounding rock under actual monitoring, is the real prediction data obtained by denormalizing the surrounding rock deformation prediction data obtained based on the ISAI-MHA-LSTM model, T n is the number of days for surrounding rock deformation prediction, MAE and MSE are the mean absolute error and mean square error of the prediction model, respectively.

[0055] The method of the present invention effectively improves the accuracy of tunnel surrounding rock deformation prediction by integrating information collaborative analysis index (ISAI) and long short-term memory network (LSTM) with attention mechanism. The present invention constructs an LSTM model integrating information theory and attention mechanism, which can dynamically focus on key time series data, thereby improving the accuracy of prediction. In addition, the method also includes a model verification step, which compares the prediction results with the measured data and uses the mean absolute error (MAE) and mean square error (MSE) to evaluate the performance of the model. The present invention not only improves the informatization and automation level of tunnel surrounding rock deformation prediction, but also provides strong technical support for tunnel construction safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A schematic diagram of a process of predicting tunnel surrounding rock deformation based on deep learning according to the present invention;

[0057] Figure 2 This is the LSTM hidden layer structure diagram;

[0058] Figure 3The prediction effect of vault settlement of ISAI-MHA-LSTM model when step=5;

[0059] Figure 4 The peripheral convergence prediction effect of the ISAI-MHA-LSTM model when step=5. DETAILED DESCRIPTION

[0060] The present invention is further described below in conjunction with specific embodiments, but the present invention is not limited to these embodiments.

[0061] like Figure 1 , a tunnel surrounding rock deformation prediction method based on deep learning, comprising the following steps:

[0062] S1. Multivariate information set data collection and processing;

[0063] S1.1 Collect deformation data of tunnel surrounding rock, including peripheral convergence deformation data and arch crown settlement data of surrounding rock;

[0064] S1.2 Collect multivariate monitoring data of tunnel surrounding rock, including but not limited to geological conditions, environmental factors, etc.;

[0065] S1.3 Perform curve regression analysis on the collected multivariate information data and use the curve shape to reflect the monitoring status of the data;

[0066] S1.4 When the curve shape shows abnormal fluctuations, immediately investigate the cause and take measures.

[0067] S2. Construction of multivariate information set;

[0068] S2.1 Based on actual projects, seven information factors, namely water content, distance to section, cohesion, burial depth ratio, compression coefficient, friction angle and density, are selected to construct the initial multivariate information factor set (X′) {x 1 , x 2 ,......,x 7};

[0069] S2.2 obtain the ISAI of each information factor;

[0070] The calculation method of ISAI is as follows:

[0071]

[0072] MIC Zi is the information factor x′ in the surrounding convergent deformation data i The calculated maximum mutual information coefficient, MIC Gi is the information factor x′ in the vault settlement data iThe calculated maximum mutual information coefficient and the results of each information factor are shown in Table 1;

[0073] Table 1

[0074] Information Factor <![CDATA[MIC Gi ]]> <![CDATA[MIC Zi ]]> ISAI Moisture content 0.777 0.951 0.21 Distance to section 0.580 0.742 0.16 Cohesion 0.564 0.705 0.15 Depth ratio 0.583 0.559 0.14 Compression factor 0.423 0.534 0.12 Friction angle 0.495 0.424 0.11 density 0.482 0.371 0.11

[0075] S2.3 Assign ISAI as a weight to each corresponding information factor, and finally form a multivariate factor information set (X) = {x 1 , x 2 ,......,x 8}.

[0076] S3. Data pre-processing;

[0077] S3.1 Data cleaning, including removing duplicate records in the data and correcting or deleting obviously erroneous data entries;

[0078] S3.2 Outlier detection and processing, including identifying outliers in the data that do not conform to the norm. These outliers may be caused by measurement errors or other abnormal conditions. The processing methods include deletion, replacement or adjustment using statistical methods;

[0079] S3.3 Data standardization, that is, converting data to a unified scale to eliminate the impact of different dimensions and magnitudes, and normalizing the data to the range of [0, 1];

[0080] Furthermore, the formula for data standardization is:

[0081]

[0082] Among them, a * is the value after data standardization, a is the original value of the data, min(G) is the minimum value of the original data set, and range(G) is the value range of the original data set.

[0083] S4. Input the data set and divide the source data into a training set and a test set;

[0084] S4.1 defines the input data set as D = {d 1 , d 2 ,...d j ,...d n};

[0085] d j ={x 1j , x 2j , x 3j , x 4j , x 5j , x 6j , x 7j , x8j}

[0086] Among them, d j represents the jth group of data;

[0087] S4.2 Divide the data after the influence factor correlation analysis and weight assignment into a training set D′ train and the test set D′ test ;

[0088] D train ={d 1 , d 2 ,…d j}

[0089] D′ test ={d j+1 , d j+2 ,...d n}

[0090] According to the step size, 80% of the data set is used as the training set D′ train , 20% as the test set D′ test .

[0091] S5.ISAI-MHA-LSTM model construction;

[0092] ISAI-MHA-LSTM model, where the LSTM hidden layer includes:

[0093] I. Construct the forget gate unit f ( t ) :f ( t ) =σ{G f [s (t-1) ,x (t) ]+b f};

[0094] II. Constructing the input gate unit i (t) :i (t) =σ{G i [s (t-1) ,x (t) ]+b i};

[0095] III. Construct output gate unit o (t) :o (t) =σ{Go[s (t-1) ,x (t) ]+b o};

[0096] The output state of the LSTM layer is s (t) =o (t) *tanh(c(t) )

[0097] c (t) =f (t) *c (t-1) +i (t) *h (t) ;

[0098] h(t)=tanh{G c [s (t-1) ,x (t) ]+b c}

[0099] Among them, G f is the weight matrix of the forget gate unit, G i is the weight matrix of the input gate unit, G o is the weight matrix of the output gate unit, b f is the bias term of the forget gate unit, b i is the bias term of the input gate unit, b o is the bias term of the output gate unit, h(t) is the temporary unit state defined by the LSTM network, tanh() is the hyperbolic tangent function, G c is the weight matrix, b c is the bias term;

[0100] The LSTM hidden layer unit structure is as follows Figure 3 shown.

[0101] The MHA layer calculates the weight matrix:

[0102] Define the weight matrix ω q ,ω v ,ω k ;

[0103] The output of each time step is linearly transformed as k and v, and the output of the last time step is multiplied by ω q As q;

[0104] Perform a scaled dot product calculation on the weight matrix to obtain the attention weight matrix W:

[0105]

[0106] where d k is the dimension of k;

[0107] After performing h times of scaled dot products, the h attention weight matrices W are concatenated. The scaled dot product is calculated as follows:

[0108]

[0109] After splicing, it passes through the linear layer W oPerform a linear transformation to obtain the final output weight matrix ω z ;

[0110] The output of the MHA layer is input into the fully connected layer.

[0111] S6. Define the LSTM layer training step LSTM(h,c,x t )=(h new ,c new ), adjust the model parameters, where h is the hidden state, c is the cell state, and x t is the input at time step t;

[0112] S6.1 sets the hidden layer size and activation function, and adds the Dropout method to the LSTM layer;

[0113] S6.2 sets the model prediction interval to 1 and the input step size to 10, based on which the sliding prediction predict(x 10:10+1 )=output 10+1 ;

[0114] S6.3 In the test data set, 7 groups of samples are selected for final prediction, and these prediction results are saved;

[0115] Based on the above steps, the calculation process of the ISAI-MHA-LSTM model is as follows:

[0116] I. Set the data set D′={d 1 , d 2 , d t ,...d s},d t ={d′ t , d′ t+1 ...d′ j-s+t-1} Input the hidden layer of LSTM;

[0117] Among them, s is the window size for time series analysis, d′ t The data at time t

[0118] II. The output of the LSTM hidden layer is O′={o 1 , o 2 , o t ,... s},o t ={o′ t , o′ t+1 ...o′ j-s+t-1};

[0119] III. Calculate the self-attention score in the MHA layer, output the weight matrix, and input the weighted data into the fully connected layer;

[0120] IV. The actual output after the fully connected layer is R = {r 1 , r 2 , r t ,...r s}

[0121] S7. Select indicators to verify and evaluate the model;

[0122] S7.1 Verify the prediction effect of the model and compare the predicted data with the measured data. The mean absolute error (MAE) and mean square error (MSE) are used to evaluate the model effect;

[0123] S7.2 obtains the real predicted data through denormalization and compares it with the measured data, such as Figure 3 , 4 ;

[0124] Furthermore, the calculation formulas for the mean absolute error (MAE) and mean square error (MSE) mentioned in S7 are as follows:

[0125]

[0126] in, is the actual deformation value of the surrounding rock under actual monitoring, is the real prediction data obtained by denormalizing the surrounding rock deformation prediction data obtained based on the ISAI-MHA-LSTM model, T n is the number of days for surrounding rock deformation prediction, MAE and MSE are the mean absolute error and mean square error of the prediction model, respectively. The specific results are shown in Table 2.

[0127] Table 2

[0128]

[0129] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-factor tunnel surrounding rock deformation prediction method based on deep learning, characterized in that: The following steps are involved: S1. Multivariate information set data collection and processing; S2. Construction of multivariate information set; S3. Data pre-processing; S4. Input the data set and divide the source data into a training set and a test set; S5.ISAI-MHA-LSTM model construction; S6. Define the LSTM layer training step size and adjust the model parameters; S7. Select indicators to verify and evaluate the model.

2. The method for predicting deformation of a multivariate tunnel surrounding rock based on deep learning according to claim 1, characterized in that: The step S1 comprises the following sub-steps: S1.1 Collect deformation data of tunnel surrounding rock, including peripheral convergence deformation data and arch crown settlement data of surrounding rock; S1.2 Collect multivariate monitoring data of tunnel surrounding rock, including geological conditions and environmental factors; S1.3 Perform curve regression analysis on the collected multivariate information data and use the curve shape to reflect the monitoring status of the data; S1.4 When the curve shape shows abnormal volatility, immediately investigate the cause and take measures.

3. The method for predicting deformation of a multivariate tunnel surrounding rock based on deep learning according to claim 2, characterized in that: The step S2 comprises the following sub-steps: S2.1 Based on actual engineering, seven information factors, namely water content, distance to section, cohesion, burial depth ratio, compression coefficient, friction angle and density, are selected to construct the initial multivariate information factor set (X′) = {x′1, x′2, ..., x′7}; S2.2 Obtain the integrated information synergy analysis index of each information factor; S2.3 assigns the integrated information collaborative analysis index as a weight to each corresponding information factor, and finally forms a multivariate factor information set (X) = {x1, x2, ..., x8} together with the deformation data; The calculation method of the integrated information collaborative analysis index mentioned in step S2.2 is as follows: MIC Zi is the information factor x′ in the surrounding convergent deformation data i The calculated maximum mutual information coefficient, MIC Gi is the information factor x′ in the vault settlement data i The maximum mutual information coefficient calculated.

4. The method for predicting deformation of a multivariate tunnel surrounding rock based on deep learning according to claim 3 is characterized in that: The step S3 comprises the following sub-steps: S3.1 Data cleaning: remove duplicate records in the data, and correct or delete obviously erroneous data entries; S3.2 Outlier detection and processing: Identify outliers in the data that do not conform to the norm. These outliers may be caused by measurement errors or other abnormal conditions. The processing methods include deletion, replacement or adjustment using statistical methods; S3.3 Data standardization, that is, converting data to a unified scale to eliminate the impact of different dimensions and magnitudes, and normalizing the data to the range of [0, 1]; Among them, the formula for data standardization mentioned in step S3.3 is: Among them, a * is the value after data standardization, a is the original value of the data, min(G) is the minimum value of the original data set, and range(G) is the value range of the original data set.

5. The method for predicting deformation of multivariate tunnel surrounding rock based on deep learning according to claim 4, characterized in that: The step S4 comprises the following sub-steps: S4.1 defines the input data set as D = {d1, d2, ...d j ,...d n }; d j ={x 1j ,x 2j ,x 3j ,x 4j ,x 5j ,x 6j ,x 7j ,x 8j } Among them, d j represents the jth group of data; S4.2 The data after the correlation analysis of influencing factors and weight assignment are divided into training set D train and the test set D′ test ; D′ train ={d1,d2,...d j } D′ test ={d j+1 ,d j+2 ,...d n } According to the step size, 80% of the data set is used as the training set D′ train , 20% as the test set D′ test。 6. The method for predicting deformation of multivariate tunnel surrounding rock based on deep learning according to claim 5, characterized in that: The step S6 comprises the following sub-steps: S6.1 sets the hidden layer size and activation function, and adds the Dropout method to the long and short-term memory network layers; S6.2 sets the model prediction interval to 1 and the input step size to 10, based on which the sliding prediction predict(x 10 : 10+1 )=output m+1 ; S6.3 In the test data set, 7 groups of samples are selected for the final prediction, and these prediction results are saved.

7. The method for predicting deformation of multivariate tunnel surrounding rock based on deep learning according to claim 6, characterized in that: The calculation process of the ISAI-MHA-LSTM model is as follows: I. Set the data set D′={d1, d2, d t ,...d s }, d t ={d′ t , d′ t+1 ...d′ j-s+t-1 } Input the hidden layer of the long short-term memory network; Among them, s is the window size for time series analysis, d′ t is the data at time t; II. The output of the hidden layer of the long short-term memory network is O′={o1, o2, o t ,... s },o t ={o′ t , o′ t+ 1...o′ j-s+t-1 }; III. Calculate the self-attention score in the MHA layer, output the weight matrix, and input the weighted data into the fully connected layer; IV. The actual output after the fully connected layer is R = {r1, r2, r t ,...r s }.

8. The method for predicting deformation of multivariate tunnel surrounding rock based on deep learning according to claim 7, characterized in that: The step S7 comprises the following sub-steps: S7.1 Verify the prediction effect of the model, compare the predicted data with the measured data, and use the mean absolute error and mean square error to evaluate the model effect; S7.2 obtains the real predicted data through denormalization and compares it with the measured data; The calculation methods of mean absolute error and mean square error are as follows: in, is the actual deformation value of the surrounding rock under actual monitoring, is the real prediction data obtained by denormalizing the surrounding rock deformation prediction data obtained based on the ISAI-MHA-LSTM model, T n The number of days for the prediction of surrounding rock deformation.

Images