Intelligent time series prediction method for thrust of shield tunneling based on hybrid deep learning

Abstract

The present invention discloses an intelligent time series prediction method for thrust of a shield tunnel based on hybrid deep learning, and belongs to the field of civil engineering construction technology. The method comprises the following steps: collecting the original data set of the project; processing the original data to construct a shield thrust prediction data set; dividing the data set into a training set, a validation set and a test set, analyzing, combining, training and verifying the features; evaluating the prediction results of the test set through an evaluation index module, and outputting an Adam‑LSTM model + an optimal feature group to achieve real-time prediction of shield thrust. The present invention adopts the above-mentioned intelligent time series prediction method for thrust of a shield tunnel based on hybrid deep learning, and can predict the shield thrust in real time through feature combination and prediction model, and is applied to actual engineering to achieve more accurate prediction of shield thrust, and provide a theoretical basis and construction guidance for thrust setting during shield excavation, which has significant engineering benefits and excellent engineering utilization value.

Description

Technical Field

[0001] The present invention relates to the technical field of civil engineering construction, and in particular to an intelligent time series prediction method for thrust of a shield tunneling excavation based on hybrid deep learning. Background Art

[0002] The shield method has become the main construction method for urban rail transit, especially for tunnel excavation in densely populated urban areas, due to its advantages of integrated excavation and support, small disturbance to the soil, short operation time, high construction efficiency, good seismic performance and safe excavation. In actual engineering, due to the complex uncertainty and spatial variability of the strata, the current setting of the thrust of the shield machine is mainly determined by engineers. As a result, the advancement of the shield project can only rely on the personal experience of engineers, and it is impossible to use existing theories and models to analyze the excavation parameters of the shield machine in real time and guide the setting of the thrust of the shield machine. Therefore, it is crucial to predict the thrust of the shield machine.

[0003] In recent years, with the development of computer technology, intelligent technologies such as machine learning and deep learning with large-scale data processing capabilities have provided new research ideas and solutions for many non-computer problems. Some research has been conducted at home and abroad on the prediction of shield machine thrust, but these studies are mainly for data analysis of completed shield projects, and it is difficult to form a relatively mature shield thrust intelligent prediction system that can be used in engineering. In addition, since shield projects will generate a large amount of engineering data, there is currently no clear method for processing and selecting data and features for artificial intelligence models.

[0004] In summary, the intelligent prediction problem of shield machine thrust still has the following difficulties: ① Shield tunneling is a continuous process. Using recent past data to predict future data can better predict shield machine thrust. This requires considering the temporal nature of data and has high prediction difficulty; ② Shield machine thrust prediction requires multiple features as input. Some features not only do not contribute to model prediction, but may also affect the model's learning of the potential laws between features, reducing prediction accuracy; ③ The original data lacks standardized preprocessing steps, and there are discrete values ​​and missing values ​​in the data set, which affects the normal prediction of the model; ④ Large differences in measurement units and values ​​will also affect the weight ratio of the features and reduce the prediction calculation efficiency of the model. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent time series prediction method for shield tunneling thrust based on hybrid deep learning. Through feature combination and prediction model, the shield thrust can be predicted in real time. It can be applied to actual engineering to achieve more accurate prediction of shield thrust, provide a theoretical basis and construction guidance for thrust setting during shield tunneling, and has significant engineering benefits and excellent engineering utilization value.

[0006] To achieve the above object, the present invention provides a shield tunneling thrust intelligent time series prediction method based on hybrid deep learning, comprising the following steps:

[0007] S1. Obtain the original data set of the project through the data acquisition module;

[0008] S2, using the data classification module, data preprocessing module and data dimension integration module to process the original data and construct a shield thrust prediction data set;

[0009] S3, divide the data set into training set, validation set and test set, analyze and combine the features through the feature combination module based on the SHAP model, and train and verify different combinations through the constructed Adam-LSTM model; among them, the transfer learning module imports the pre-trained network weights into the model;

[0010] S4. Evaluate the prediction results of the test set through the evaluation index module, obtain the network connection weights, bias and corresponding optimal feature combination of the Adam-LSTM model, and output the Adam-LSTM model + optimal feature group to achieve real-time prediction of shield thrust.

[0011] Preferably, the data acquisition module is composed of equipment related to shield tunneling engineering, specifically including earth pressure balance shield machine, formation condition collection instrument, and tunnel design equipment; the parameters collected from the shield machine include relevant parameters of the propulsion system, tunneling system, soil discharge system, gradual system and control system, where the collection unit is each ring of shield lining segments.

[0012] Preferably, the data classification module performs information annotation and classification on the shield engineering data contained in the original data set; wherein the information annotation includes the ring number of the shield tunneling, and the corresponding geomechanical parameters, tunnel geometric parameters, and shield tunneling parameters of the current ring; the annotated and classified data is saved in a file in .xml format.

[0013] Preferably, the data preprocessing module reads the .xml information file corresponding to the current shield tunneling project and preprocesses the data part;

[0014] First, scan the original data to identify normal values, discrete values, and missing values ​​in the data;

[0015] Secondly, the 3σ criterion is used to identify and correct discrete values, and the same type of mean interpolation is used to fill in missing values.

[0016] Preferably, the data dimension integration module reads the preprocessed data, calculates the mean and standard deviation of the feature parameters, uses Z-score standardization, performs linear transformation on the data set, and reduces the dimension of the data, as shown below:

[0017]

[0018]

[0019]

[0020] Among them, the mean of the standardized data set after dimensionality reduction is 0, the standard deviation is 1, and the range is [-1,1]; μ (j) and σ (j) represents the mean and standard deviation of feature j, is the original data of feature j, n represents the number of parameters of feature j, and represents the number of rings involved in thrust prediction of shield tunneling. The data obtained after the feature j parameter is standardized.

[0021] Preferably, the core of the feature combination module based on the SHAP model is to use the interpretable machine learning model SHAP to evaluate the contribution of features to model prediction, and combine features based on this as a judgment criterion;

[0022] The feature combination module based on the SHAP model reads the feature information and related parameters in the shield thrust prediction database, calculates the Shapleyvalue of each feature and arranges the features in descending order based on this; the feature combination starts with selecting the feature with the highest contribution, and then adds a feature of the next order in turn to form a new combination; let I = {1, 2, ... n} be a set of n features, and the calculation method of Shapleyvalue is as follows:

[0023]

[0024]

[0025] Among them, x i is the Shapley value of feature i; S i is the set formed by all subsets containing feature i in I, s is one of the subsets; |s| is the number of features in subset s; v(s) represents the influence of subset s on the prediction result, and v(s\i) represents the influence of the original subset s without feature i on the prediction result; [v(s)-v(s\i)] represents the contribution of feature i in combination s to the prediction result, that is, the marginal contribution of feature i; ω(|s|) represents the weight corresponding to the contribution.

[0026] Preferably, the Adam-LSTM model is composed of an adaptive moment estimation optimized Adam and a long short-term memory recurrent neural network LSTM; the core of the model is LSTM, which is based on the recurrent neural network RNN ​​and is equipped with a gating unit to realize the network's forgetting, updating, storage and output of information;

[0027] (1) The forget gate is based on the output h of the previous cell. t-1 and the current input x t Through σ, a Sigmoid nonlinear mapping is performed to obtain a vector f with a value of each dimension between 0 and 1. t , as shown below:

[0028] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0029] (2) The input gate is composed of the output h of the previous cell. t-1 and the current input x t Through a Sigmoid nonlinear mapping, and then through a tanh function layer to generate new candidate information a t , as shown below:

[0030] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0031] a t =σ(W a ·[h t-1 ,x t ]+b a )

[0032] (3) Cell old state C t-1 and f t Multiply to discard the information that is determined to be forgotten, and add i t ⊙a t As new information is added to the cell state, the cell state is updated as follows:

[0033] C t =f t ⊙C t-1 +i t ⊙a t

[0034] (4) The output gate is determined by the output h of the previous cell. t-1 and the current input x tThrough a Sigmoid function layer, the cell state C t After being processed by a tanh function layer and multiplied by the output of the Sigmoid layer, the part to be output is finally determined and output as shown below:

[0035] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0036] h t =o t ⊙tanh(C t )

[0037] Among them, W represents weight, b represents bias, and the subscripts f, i, a, and o correspond to the output of the formula respectively. The model achieves reasonable prediction by continuously adjusting weights and biases.

[0038] Preferably, the adaptive moment estimation optimization, i.e., Adam optimization, improves the gradient and learning rate based on the stochastic gradient descent method as follows:

[0039]

[0040] Among them, θ t+1 and θ t Respectively represent the loss function values ​​at time i+1 and i, ε is a very small number; represents the learning rate of Adam optimization, η represents the step size of each iteration, The accumulation of past gradients introduced into the denominator of the learning rate; Represents the gradient of Adam optimization;

[0041] The calculation formulas for learning rate and gradient are as follows:

[0042] m t =β1·m t-1 +(1-β1)·g t

[0043]

[0044] β1=0.9, β2=0.999, m0=0, γ0=0

[0045]

[0046]

[0047] Among them, m t represents the weighted moving average of the gradient, γt Represents the weighted moving average of the learning rate. The initial values ​​m0 and γ0 are 0. After bias correction is performed on m0 and γ0, it becomes and When t>30, When t>3000,

[0048] Preferably, the transfer learning module sets an approximate network architecture parameter trained based on other shield engineering data sets into the Adam-LSTM model, including the connection weights and biases between each hidden layer; the Adam-LSTM model obtains prior knowledge and conducts further learning on the specific project.

[0049] Preferably, the evaluation index module reads the trained model and evaluates and judges the training results;

[0050] The mean square error (MSE) is used as the fitness function of the validation process, as shown below:

[0051]

[0052] in, is the predicted value of the thrust of the i-th ring in the training set; y i is the true value of the thrust of the i-th ring in the training set;

[0053] The coefficient of determination R 2 The evaluation indicators of the test results are as follows:

[0054]

[0055] in, is the predicted value of the thrust of the i-th ring in the test set; y i is the true value of the thrust of the i-th ring in the test set; The error in predicting the values ​​for the test set; is the error generated by the mean of the test set, and y is the average thrust of ring i from 1 to n in the test set.

[0056] Therefore, the present invention adopts the above-mentioned shield tunneling thrust intelligent time series prediction method based on hybrid deep learning, and the beneficial effects are as follows:

[0057] 1. For the original data, the 3σ criterion is used to process discrete values, and the same type of mean interpolation is used to fill the missing values; the Z-score standardization is used to reduce the dimension of the original data, which provides suitable input parameters for the shield thrust prediction model and improves the calculation efficiency and prediction accuracy of the model;

[0058] 2. Compared with the existing shield thrust prediction methods, the different degrees of influence of the timing and characteristics of shield tunneling on the shield thrust prediction are taken into account, and the problem is addressed from two aspects: the shield tunneling mechanism and the data correlation. The network properties of LSTM are used to consider the timing of shield thrust from the input parameter level; the SHAP model is used to quantitatively analyze the contribution of different characteristic parameters to model prediction, and a specific feature combination method is proposed, which enables the present invention to achieve a more accurate prediction of shield thrust at the level of controlling model input parameters;

[0059] 3. After training, the final feature combination and prediction model can predict the thrust of the shield in real time, can be used in actual projects, and provide a theoretical basis and construction guidance for the thrust setting during shield excavation, which has significant engineering benefits and excellent engineering utilization value.

[0060] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the framework of the intelligent time series prediction method of thrust of shield tunneling based on hybrid deep learning of the present invention;

[0062] Figure 2 It is a flow chart of raw data processing of the intelligent time series prediction method for thrust of shield tunneling based on hybrid deep learning of the present invention;

[0063] Figure 3 The present invention is a flowchart of training data processing for the intelligent time series prediction method for thrust of shield tunneling based on hybrid deep learning. DETAILED DESCRIPTION

[0064] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0065] The intelligent time series prediction method of thrust of shield tunneling based on hybrid deep learning includes the following steps:

[0066] S1. Obtain the original data set of the project through the data acquisition module;

[0067] S2, using the data classification module, data preprocessing module and data dimension integration module to process the original data and construct a shield thrust prediction data set;

[0068] S3, divide the data set into training set, validation set and test set, analyze and combine the features through the feature combination module based on the SHAP model, and train and verify different combinations through the constructed Adam-LSTM model; among them, the transfer learning module imports the pre-trained network weights into the model;

[0069] S4. The prediction results of the test set are evaluated through the evaluation index module to obtain the network connection weights, biases and corresponding optimal feature combinations of the Adam-LSTM model, and the Adam-LSTM model + optimal feature group are output. The output model can realize the real-time prediction of shield thrust, assist and guide the setting of shield thrust in shield excavation projects, and realize real-time prediction of shield thrust.

[0070] Example

[0071] like Figure 1 As shown in the figure, the system framework of the intelligent time series prediction method for thrust of shield tunneling based on hybrid deep learning includes data acquisition module, data classification module, data preprocessing module, data dimension integration module, feature combination module based on SHAP model, transfer learning module, Adam-LSTM model, evaluation index module and Adam-LSTM model + optimal feature combination.

[0072] The data acquisition module is composed of equipment related to shield tunneling engineering, such as earth pressure balance shield machine, ground condition acquisition instrument, tunnel design equipment, etc. Relevant parameters are collected from the shield machine's propulsion system, tunneling system, soil discharge system, gradual system, and control system, and the collection unit is each ring of shield lining segments.

[0073] The function of the data classification module is to annotate and classify the shield engineering data contained in the original data set. The information annotation includes the ring number of each shield tunneling ring, as well as the corresponding geomechanical parameters, tunnel geometry parameters, and shield tunneling parameters of the current ring. The annotated and classified data is saved in a .xml format file. The .xml file is <filename> ， <path> ， <project>The information segments include <project>Also included <type> ， <label>， <number>The specific information used to predict the thrust of the shield machine is as follows:

[0074]

[0075] The first value corresponds to the lining ring number of the shield tunnel project, which can correspond to the coordinate information of the specific excavation process of the shield machine, and also serves as the label of the current ring lining. The following values ​​are three different types of parameters in the excavation process of each ring, corresponding to the geomechanical parameters, tunnel geometry parameters, and shield excavation parameters of the current ring.

[0076] Geomechanical parameters must include: stratum soil density, equivalent internal friction angle of soil, soil deformation modulus, soil Poisson's ratio, soil permeability coefficient and other parameters; tunnel geometric parameters must include: tunnel vault burial depth, tunnel radius and other parameters; shield excavation parameters must include: thrust, cutter head torque, excavation speed, soil bin pressure, excavation volume, synchronous grouting volume, synchronous grouting pressure and other parameters.

[0077] The data set is classified into three categories and required to collect important data from the three types in order to ensure the comprehensiveness of data collection, thereby enhancing the accuracy of model prediction, improving the generalization ability of the model, and reducing prediction deviation and overfitting.

[0078] The data preprocessing module reads the .xml information file corresponding to the current shield tunneling project and preprocesses the data. The specific process includes: scanning the original data, distinguishing the normal values, discrete values ​​and missing values ​​in the data. The 3σ criterion is used to identify and correct discrete values, and the same type of mean interpolation is used to fill in missing values.

[0079] The shield thrust prediction dataset takes into account many types of data. Not only may these data have huge differences in numerical expression due to different measurement units in different projects, but the values ​​of different types of characteristic parameters often differ by orders of magnitude.

[0080] The data dimension integration module reads the preprocessed data and calculates the mean and standard deviation of the characteristic parameters according to the characteristics of the shield thrust prediction data set. The data is standardized with the help of the mean and standard deviation, and the data set is linearly transformed according to its type using Z-score standardization. All types of data are reduced in dimension so that their values ​​are not affected by the measurement unit and become standardized data sets with a mean of 0, a standard deviation of 1, and a range of [-1,1], thereby improving the comparability of the data. Among them, the calculation formulas for the mean, standard deviation, and characteristic parameter standardization are as follows:

[0081]

[0082]

[0083]

[0084] Among them, μ (j) and σ (j) represents the mean and standard deviation of feature j, is the original data of feature j, n represents the number of parameters of feature j, and represents the number of rings involved in thrust prediction of shield tunneling. The data obtained after the feature j parameter is standardized.

[0085] The data after data dimension integration will replace the original data, and the corresponding specific information of predicted shield machine thrust will be reintegrated into the original .xml format file to form a database for shield thrust prediction. The database is divided by random selection, and 70% of the data is selected as the training data set, 20% of the data is selected as the verification data set, and 10% of the data is selected as the test data set. Among them, the training set contains data information for network training for Adam-LSTM model training and learning; the verification set contains data information for network verification to verify whether the training of the Adam-LSTM model has enabled the model to reach the level that can be used to predict shield thrust; the test set contains data information for network testing to compare which set of feature combinations can train the Adam-LSTM model network connection weights and biases to achieve the best results in shield thrust prediction.

[0086] The shield thrust prediction dataset contains a wide variety of different features, but not all features can help the model predict thrust. Some features may not only have no positive effect on model prediction, but may also affect the model's learning of the potential rules between features, reducing the model's prediction efficiency and accuracy. Therefore, screening and effective combination of features is an essential and effective approach in the intelligent time series prediction method of shield thrust.

[0087] The core of the feature combination module based on the SHAP model is to use an interpretable machine learning model (SHapleyAdditive exPlanations, referred to as the SHAP model) to evaluate the contribution of features to model predictions, and use this as a criterion for combining features. In cooperative game theory, each individual wants to maximize his or her own results. By calculating the contribution of individuals in cooperation, we can be sure of the importance of the individual. The SHAP model is an additive explanatory model built under the inspiration of cooperative game theory. It regards all features as "contributors" to model predictions. For each predicted sample, the model will generate a predicted value. The SHAP model will quantify the degree to which each feature in the sample contributes to the model's prediction value, namely the Shapley value.

[0088] The feature combination module based on the SHAP model reads the feature information and related parameters in the shield thrust prediction database, calculates the Shapley value of each feature and arranges the features in descending order based on this. The feature combination starts with selecting the feature with the highest contribution, and then adds a feature of the next order to form a new combination. The calculation formula of the Shapley value is as follows:

[0089]

[0090]

[0091] Among them, x i is the Shapley value of feature i; S i is the set formed by all subsets of I containing feature i, I = {1, 2, ... n} is a set of n features; s is one of the subsets; |s| is the number of features in subset s; v(s) represents the influence of subset s on the prediction result, and v(s\i) represents the influence of the original subset s without feature i on the prediction result; [v(s)-v(s\i)] represents the "contribution" of feature i in combination s to the prediction result, that is, the marginal contribution of feature i; ω(|s|) represents the weight corresponding to the "contribution".

[0092] Due to the continuity of shield tunneling construction and the complexity of the tunneling stratum environment, the thrust of the shield machine is a time-series parameter, and the thrust of the current ring is affected by the parameters related to the number of excavated rings. Therefore, a model that can consider the time-series of the thrust of the shield machine can more accurately predict the thrust of the shield machine.

[0093] The Adam-LSTM model is a neural network that is mixed with an optimization algorithm and is used to process time-series data. Its core is the long short-term memory recurrent neural network, or LSTM. LSTM is based on the recurrent neural network RNN. By adding the corresponding function layer as a gating unit, the network can forget, update, store and output information.

[0094] LSTM consists of four interactive parts: forget gate, input gate, cell state, and output gate. The cell state is used as a memory unit, and the input gate, output gate, and forget gate are used to update the information in the memory unit. Due to the existence of the memory unit and the gate mechanism, LSTM can effectively merge the input information and the information of the hidden state layer, thereby realizing the ability to learn long-term information.

[0095] The forget gate is used to control how much past information will be selected and forgotten, which is determined by the output h of the previous cell. t-1 and the current input x t Through σ, a Sigmoid nonlinear mapping is performed to obtain a vector f with a value of each dimension between 0 and 1. t , the formula is as follows:

[0096] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0097] Among them, W represents weight, b represents bias, and the subscripts f, i, a, and o correspond to the output of the formula respectively. The model continuously adjusts the weight and bias to achieve the final reasonable prediction; the forget gate vector will be combined with the cell state C t-1 When the value of the cell state is multiplied by 0, the result is close to 0, indicating that the information is forgotten; when it is multiplied by 1, the value itself is obtained, indicating that the information is completely stored. Through this mechanism, the forget gate realizes the memory of important information and the forgetting of unimportant information.

[0098] The input gate is used to control the storage ratio of the current input information, which is determined by the output h of the previous cell. t-1 and the current input x t Through a Sigmoid nonlinear mapping, the function layer determines which part of the cell state value to update, and then through a tanh function layer to generate new candidate information a t , the formula is as follows:

[0099] i t =σ(W i ·[h t-1 ,x t ]+b i )

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

[0101] The update formula of the cell state is as follows:

[0102] C t =f t ⊙C t-1 +i t ⊙a t

[0103] Among them, the old state of the cell C t-1 and f t Multiply to discard the information that is determined to be forgotten, and add i t ⊙a t As new information is added to the cell state, the cell state is updated.

[0104] The function of the output gate is to determine how much current information to output, based on the output h of the previous cell. t-1 and the current input x t A Sigmoid function layer is used to determine which parts of the cell state to output, and then the cell state C t After being processed by a tanh function layer and multiplied by the output of the Sigmoid layer, the part to be output is finally determined and output. The formula is as follows:

[0105] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0106] h t =o t ⊙tanh(C t )

[0107] Adaptive Moment Estimation (Adam optimization) is an optimization algorithm that can optimize random numbers in LSTM models. Adam optimization improves the gradient and learning rate based on the stochastic gradient descent method, as shown below:

[0108]

[0109] Among them, θ t+1 and θ t Represent the loss function values ​​at time i+1 and i respectively. ε is a very small number to prevent the denominator from being 0. Generally, ε=0.0001; represents the learning rate of Adam optimization, η represents the step size of each iteration, The denominator of the learning rate introduces a cumulative sum of past gradients. This change makes the denominator of the learning rate larger in the section with a large gradient, and the overall step size smaller, to avoid excessive step size, running out of the minimum point or oscillating near the minimum point; in the section with a large gradient, the denominator becomes smaller, and the overall step size becomes larger, which accelerates the learning rate to cross the local flat area, improves the convergence speed, and realizes adaptive adjustment of the learning rate. Represents the gradient of Adam optimization.

[0110] The calculation formulas for learning rate and gradient are as follows:

[0111] m t =β1·m t-1 +(1-β1)·g t

[0112]

[0113] β1=0.9, β2=0.999, m0=0, γ0=0

[0114]

[0115]

[0116] Among them, m t represents the weighted moving average of the gradient, γ t represents the weighted moving average of the learning rate; since the initial values ​​of m0 and γ0 are 0, their initial changes are slow. In order to accelerate the initial changes of the gradient and learning rate, m0 and γ0 are biased and become and When t>30, When t>3000,

[0117] The transfer learning module sets the parameters of an approximate network architecture trained based on other shield engineering data sets into a new Adam-LSTM model. The network architecture parameters include the connection weights and biases between each hidden layer. In this way, the model can obtain some prior knowledge and further learn about specific projects on this basis, thereby saving the training cost of the model and improving the training speed and learning efficiency of the model.

[0118] The features combined by the feature combination module based on the SHAP model are input into the Adam-LSTM model to train the prediction of the shield thrust. Considering the time series of predicting the shield thrust, when predicting the shield thrust of the i-th ring, the relevant parameters of the i-10, i-9, ..., i-2, i-1 rings, that is, the ten rings before the current ring, are used as the input parameters for predicting the current ring shield thrust.

[0119] The evaluation index module will read the trained model and the corresponding validation set and test set, and evaluate and judge the training results. The trained model uses the inverse standardized validation set to verify the prediction results, and the mean square error MSE is used as the fitness function of the verification process, as shown below:

[0120]

[0121] in, is the predicted value of the thrust of the i-th ring in the training set; y i is the true value of the thrust of the i-th ring in the training set.

[0122] When the fitness function of a set of feature combinations after training does not tend to converge, it is considered that the training of this set has not ended. The model will train the network connection weights, biases and other parameters of the model based on Adam optimization to reduce the fitness function value of the validation set prediction results; when the fitness function of a set of feature combinations after training tends to converge, it is considered that the training of this set is over. The model prediction results are tested 20 times using the inverse standardized test set, and the determination coefficient R of the 20 prediction results is used. 2 The evaluation indicators of the test results are as follows:

[0123]

[0124] in, is the predicted value of the thrust of the i-th ring in the test set; y i is the true value of the thrust of the i-th ring in the test set; The error in predicting the values ​​for the test set; is the error generated by the mean of the test set, and y is the average thrust of ring i from 1 to n in the test set. 2 It can be used to measure the prediction effect of the model and can intuitively express the advantages and disadvantages of the model. It is a relatively clear evaluation standard.

[0125] After testing all feature combinations, in the prediction results of the test set, R 2 The largest feature combination and its corresponding network connection weights, bias and other parameters can be considered as the optimal shield thrust intelligent time series prediction combination and parameters adapted to the Adam-LSTM model, and are output to the Adam-LSTM model + optimal feature combination.

[0126] For shield tunneling projects that are underway, as long as the relevant parameters of the ten rings before the current ring can be collected, they can be used as input parameters, and the Adam-LSTM model + optimal feature combination can be used to predict the shield thrust in real time, thus realizing real-time prediction of the shield thrust as the project progresses.

[0127] Therefore, the present invention adopts the above-mentioned intelligent time series prediction method of shield tunneling thrust based on hybrid deep learning. Through feature combination and prediction model, the shield thrust can be predicted in real time, and applied to actual engineering projects to achieve more accurate prediction of shield thrust, providing a theoretical basis and construction guidance for thrust setting during shield tunneling, which has significant engineering benefits and excellent engineering utilization value.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.< / number> < / label> < / type> < / project> < / project> < / path> < / filename>

Claims

1. An intelligent time series prediction method for thrust of shield tunneling based on hybrid deep learning, characterized in that: The following steps are involved: S1. Obtain the original data set of the project through the data acquisition module; S2, using the data classification module, data preprocessing module and data dimension integration module to process the original data and construct a shield thrust prediction data set; The data classification module annotates and classifies the shield engineering data contained in the original data set; the information annotated includes the ring number of the shield tunneling, as well as the corresponding geomechanical parameters of the current ring, tunnel geometric parameters, and shield tunneling parameters; the annotated and classified data is saved in a .xml format file; S3, divide the data set into training set, validation set and test set, analyze and combine the features through the feature combination module based on the SHAP model, and train and verify different combinations through the constructed Adam-LSTM model; among them, the transfer learning module imports the pre-trained network weights into the model; The core of the feature combination module based on the SHAP model is to use the interpretable machine learning model SHAP to evaluate the contribution of features to model prediction, and use this as the judgment standard to combine features; The feature combination module based on the SHAP model reads the feature information and related parameters in the shield thrust prediction database, calculates the Shapleyvalue of each feature and arranges the features in descending order based on this; the feature combination starts with selecting the feature with the highest contribution, and then adds a feature of the next order in turn to form a new combination; let I = {1, 2, ... n} be a set of n features, and the calculation method of Shapleyvalue is as follows: Among them, x i is the Shapley value of feature i; S i is the set formed by all subsets containing feature i in I, s is one of the subsets; |s| is the number of features in subset s; v(s) represents the influence of subset s on the prediction result, and v(s\i) represents the influence of the original subset s after removing feature i on the prediction result; [v(s)-v(s\i)] represents the contribution of feature i in combination s to the prediction result, that is, the marginal contribution of feature i; ω(s) represents the weight corresponding to the contribution; The transfer learning module sets the parameters of an approximate network architecture trained based on other shield engineering data sets into the Adam-LSTM model, including the connection weights and biases between each hidden layer; the Adam-LSTM model obtains prior knowledge and conducts further learning on specific projects; S4. Evaluate the prediction results of the test set through the evaluation index module, obtain the network connection weights, bias and corresponding optimal feature combination of the Adam-LSTM model, and output the Adam-LSTM model + optimal feature group to achieve real-time prediction of shield thrust.

2. According to claim 1, the method for intelligent time series prediction of thrust of shield tunneling based on hybrid deep learning is characterized in that: The data acquisition module is composed of equipment related to shield tunneling projects, including earth pressure balance shield machines, ground condition acquisition instruments, and tunnel design equipment; the parameters collected from the shield machines include relevant parameters of the propulsion system, tunneling system, soil discharge system, gradual system, and control system, where the acquisition unit is each ring of shield lining segments.

3. The method for intelligent time series prediction of thrust of shield tunneling based on hybrid deep learning according to claim 1 is characterized in that: The data preprocessing module reads the .xml information file corresponding to the current shield tunneling project and preprocesses the data part; First, scan the original data to identify normal values, discrete values, and missing values ​​in the data; Secondly, the 3σ criterion is used to identify and correct discrete values, and the same type of mean interpolation is used to fill in missing values.

4. The method for intelligent time series prediction of thrust of shield tunneling based on hybrid deep learning according to claim 1 is characterized in that: The data dimension integration module reads the preprocessed data, calculates the mean and standard deviation of the feature parameters, uses Z-score standardization, performs linear transformation on the data set, and reduces the data dimension, as shown below: Among them, the mean of the standardized data set after dimensionality reduction is 0, the standard deviation is 1, and the range is [-1,1]; μ (j) and σ (j) represents the mean and standard deviation of feature j, is the original data of feature j, n represents the number of parameters of feature j, and represents the number of rings involved in thrust prediction of shield tunneling. The data obtained after the feature j parameter is standardized.

5. The method for intelligent time series prediction of thrust of shield tunneling based on hybrid deep learning according to claim 1 is characterized in that: The Adam-LSTM model consists of an adaptive moment estimation optimization Adam and a long short-term memory recurrent neural network LSTM. The core of the model is LSTM, which is built on the basis of the recurrent neural network RNN ​​and has a gating unit to achieve the network's forgetting, updating, storage and output of information. (1) The forget gate is based on the output h of the previous cell. t-1 and the current input x t Through σ, a Sigmoid nonlinear mapping is performed to obtain a vector f with a value of each dimension between 0 and 1. t , as shown below: f t =σ(W f ·[h t-1 ,x t ]+b f ) (2) The input gate is composed of the output h of the previous cell. t-1 and the current input x t Through a Sigmoid nonlinear mapping, and then through a tanh function layer to generate new candidate information a t , as shown below: i t =σ(W i ·[h t-1 ,x t ]+b i ) a t =σ(W a ·[h t-1 ,x t ]+b a ) (3) Cell old state C t-1 and f t Multiply to discard the information that is determined to be forgotten, and add i t ⊙a t As new information is added to the cell state, the cell state is updated as follows: C t =f t ⊙C t-1 +i t ⊙a t (4) The output gate is determined by the output h of the previous cell. t-1 and the current input x t Through a Sigmoid function layer, the cell state C t After being processed by a tanh function layer and multiplied by the output of the Sigmoid layer, the part to be output is finally determined and output as shown below: the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ⊙tanh(C t ) Among them, W represents weight, b represents bias, and the subscripts f, i, a, and o correspond to the output of the formula respectively. The model achieves reasonable prediction by continuously adjusting weights and biases.

6. The method for intelligent time series prediction of thrust of shield tunneling based on hybrid deep learning according to claim 1 is characterized in that: Adaptive moment estimation optimization, also known as Adam optimization, improves the gradient and learning rate based on the stochastic gradient descent method as follows: Among them, θ t+1 and θ t Represent the loss function values ​​at time i+1 and i respectively, and ε is a very small number; represents the learning rate of Adam optimization, η represents the step size of each iteration, The accumulation of past gradients introduced into the denominator of the learning rate; Represents the gradient of Adam optimization; The calculation formulas for learning rate and gradient are as follows: m t =β1·m t-1 +(1-β1)·g t β1=0.9, β2=0.999, m0=0, γ0=0 Among them, m t represents the weighted moving average of the gradient, γ t Represents the weighted moving average of the learning rate. The initial values ​​m0 and γ0 are 0. After bias correction is performed on m0 and γ0, it becomes and When t>30, When t>3000, 7. The method for intelligent time series prediction of thrust of shield tunneling based on hybrid deep learning according to claim 1 is characterized in that: The evaluation index module reads the trained model and evaluates and judges the training results; The mean square error (MSE) is used as the fitness function of the validation process, as shown below: in, is the predicted value of the thrust of the i-th ring in the training set; y i is the true value of the thrust of the i-th ring in the training set; The coefficient of determination R 2 The evaluation indicators of the test results are as follows: in, is the predicted value of the thrust of the i-th ring in the test set; y i is the true value of the thrust of the i-th ring in the test set; The error in predicting the values ​​for the test set; is the error generated by the mean of the test set, is the average thrust of ring i from 1 to n in the test set.

Images