Shield segment floating lifting prediction system and method based on improved hybrid neural network

Abstract

The application discloses a shield segment floating lifting prediction system and method based on an improved mixed neural network, and belongs to the technical field of shield tunnel construction monitoring. The system comprises data acquisition, processing, prediction, fusion and interpretation output modules. The acquisition module acquires construction and structural response parameters in real time. The processing module forms input samples through abnormal value elimination, normalization and principal component analysis. In the prediction module, the XGBoost sub-model uses an improved sparrow search algorithm to optimize hyperparameters and outputs a first prediction value, and the GRU-SVM sub-model captures time sequence characteristics and outputs a second prediction value. The fusion module obtains a final floating lifting prediction value by using an error reciprocal weighting method. The interpretation output module calculates the marginal contribution value of each parameter by using a SHAP algorithm. The application considers the prediction accuracy, timeliness and interpretability, provides a scientific basis for construction regulation and control, and is suitable for multi-condition shield segment floating lifting monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of shield tunnel construction monitoring, in particular to a shield segment floating prediction system and method based on an improved hybrid neural network. BACKGROUND

[0002] The segment lining structure of a shield tunnel is the core of the forming support system, and its installation quality directly affects the stability and service life of the tunnel structure. However, during the shield propulsion process, due to the coupling effect of multiple factors such as complex geological conditions, stratum disturbance, uneven synchronous grouting of the shield tail, construction parameter fluctuations, etc., the segment often experiences a "floating" phenomenon. Floating refers to the phenomenon of arching, lifting, and even mispositioning of the segment in an unconsolidated state. Segment floating changes the assembly accuracy of the segment, causing lining leakage, cracks, misalignment, and even deformation, which poses a safety hazard.

[0003] There are two methods for predicting the impact of segment floating in the prior art. One is a theoretical prediction method based on mechanical models, such as the elastic beam model, the contact-friction model, and the shield tail gap model. The other is a monitoring-inversion method based on historical experience or numerical simulation, such as three-dimensional finite element analysis and discrete element simulation. Both methods have drawbacks. First, theoretical mechanical models are oversimplified. To enable solution, they often idealize the stratum, grout, and shield tail gap, resulting in ambiguous model representation. Second, numerical simulation methods are highly accurate but time-consuming to model and calculate, and they are overly dependent on boundary conditions, with weak timeliness. In addition, traditional machine learning methods also have the problems of insufficient accuracy and lack of interpretability, which limits the credibility of the prediction and the reference value of the decision. SUMMARY

[0004] The present application provides a shield segment floating prediction system based on an improved hybrid neural network, which solves the problems existing in the prediction of the impact of segment floating from the dimensions of precision, strong timeliness, and interpretability.

[0005] The technical solution of the present application is as follows:

[0006] A shield segment floating prediction system based on an improved hybrid neural network, comprising:

[0007] A data acquisition module that acquires construction parameters and structural response parameters in real time based on a shield machine data acquisition system and synchronous monitoring equipment;

[0008] A data processing module that performs outlier rejection, normalization, and principal component analysis operations on the construction parameters and structural response parameters to form input samples, with each parameter in the input samples being recorded as a feature input vector;

[0009] The prediction module includes an XGBoost sub-model and a GRU-SVM sub-model.

[0010] The XGBoost sub-model takes the input sample as input and outputs the first boost prediction value;

[0011] The GRU-SVM sub-model takes a feature input vector with temporal characteristics as input and outputs a second floating prediction value.

[0012] The fusion module is used to fuse the first buoyancy prediction value and the second buoyancy prediction value into a buoyancy prediction value;

[0013] An interpretation output module calculates the marginal contribution value of each feature input vector in the input sample based on the buoyancy prediction value. The marginal contribution value is used to characterize the buoyancy of the tunnel segment under the influence of construction parameters.

[0014] Furthermore, in the prediction module, the root mean square error of the XGBoost sub-model's prediction on the training set is used as the fitness function. An improved sparrow search algorithm is used to search for the hyperparameter combination that minimizes the root mean square error of the prediction as the optimal hyperparameter. Based on the optimal hyperparameter, the optimal XGBoost sub-model is constructed and used to output the first boost prediction value.

[0015] Furthermore, in the prediction module, the GRU-SVM sub-model is constructed as follows:

[0016] The state vector output by the GRU neural network is used as input to the support vector regression model for nonlinear fitting to output the second buoyancy prediction value.

[0017] Secondly, this application also provides a method for predicting the buoyancy of tunnel segments based on an improved hybrid neural network, comprising the following steps:

[0018] S1. Collect construction parameters and structural response parameters;

[0019] S2. Perform outlier removal, normalization, and principal component analysis on the collected data to establish input samples containing feature input vectors;

[0020] S3. Based on the data collected in step S2, construct the training input samples of the XGBoost sub-model, input the training input samples into the XGBoost sub-model for training, use the root mean square error of prediction as the fitness function during the training process, use the improved sparrow search algorithm to search for the hyperparameter combination that minimizes the root mean square error of prediction as the optimal hyperparameter, construct the optimal XGBoost sub-model based on the optimal hyperparameter and use it to output the first boost prediction value.

[0021] S4. Based on the data collected in step S2, construct training input samples with time-series characteristics for the GRU-SVM sub-model. Input the training input samples into the output state vector of the GRU neural network. Input the state vector into the support vector regression model for training. Obtain the fitting function by nonlinear fitting through kernel function mapping. Use the fitting result as the second floating prediction value.

[0022] S5. The first and second buoyancy prediction values ​​are fused using the error reciprocal weighting method to obtain the buoyancy prediction value;

[0023] S6. The SHAP interpretation algorithm is used to construct a linear weighted model of the input samples and their corresponding buoyancy prediction values. A baseline prediction value is set in the linear weighted model, and the marginal contribution value corresponding to each construction parameter in the input samples is calculated.

[0024] Furthermore, in step S2, 3 Outlier removal is performed according to the following principles:

[0025] Calculate the mean of each type of data collection and standard deviation It will fall within the interval Data outside the range is treated as outliers and removed; for data crossing stratigraphic fault zones and stratigraphic boundaries, if the collected data is missing, interpolation is used to supplement the data based on data collected at adjacent times.

[0026] In step S2, the collected data is normalized using the Min-Max normalization method.

[0027] Further, in step S2, the covariance matrix of the matrix formed by the normalized collected data is calculated and eigenvalue decomposition and sorting are performed. The principal component variables with a cumulative variance contribution rate greater than 95% are retained as comprehensive variables, and the comprehensive variables are used as input samples for the XGBoost sub-model.

[0028] Add time-series windows to the input samples of the XGBoost sub-model to serve as input samples for the GRU-SVM sub-model.

[0029] Further, in step S5, the root mean square error of the predictions of the XGBoost sub-model and the GRU-SVM sub-model on the validation set is first calculated. and Then calculate the predicted buoyancy value according to the following formula:

[0030] ;

[0031] In the formula, This represents the predicted rise value. This represents the first predicted rise value. This represents the second predicted rise value.

[0032] Further, in step S6, the linear weighted model is:

[0033] ;

[0034] In the formula, Indicates the baseline forecast value. Indicates the first i The marginal contribution of each construction parameter to the predicted buoyancy value.

[0035] Due to the adoption of the above technical solution, the beneficial effects of this application are as follows:

[0036] 1. This application utilizes the ISSA algorithm to globally optimize the hyperparameters of the XGBoost model, enhancing its learning ability for static features. Simultaneously, this application employs a GRU-SVM structure to characterize the dynamic evolution trend of construction parameters, enabling the model to balance time dependence and nonlinear fitting capabilities. The ISSA algorithm simulates the foraging behavior of a sparrow population, achieving global optimization of the XGBoost parameter space through the iterative evolution of individuals such as "discoverers," "followers," and "watchdogs" in the search space. The GRU neural network is an optimized recurrent neural network (RNN) structure that effectively captures the dynamic changes of construction variables over time. GRU learns the influence of historical states on current buoyancy changes through a series of gating mechanisms, and combined with SVM based on structural risk minimization theory, it achieves buoyancy value regression prediction in high-dimensional space, thus exhibiting good generalization ability.

[0037] 2. This application adaptively integrates the output results of multiple sub-models through a difference reciprocal weighting strategy, significantly reducing the model's sensitivity to outlier data or single variable failure. It exhibits high consistency and transferability under different geological formations and working conditions, meeting the needs for buoyancy risk monitoring during multi-regional, multi-stage shield tunneling construction.

[0038] 3. This application overcomes the limitations of black box models in terms of their inability to explain, and for the first time introduces the SHAP interpretability mechanism in the field of shield tunneling buoyancy prediction. It can perform principal component decomposition on each buoyancy prediction result, and clearly give the contribution of variables such as propulsion rate, grouting pressure, and shield tail gap to the current buoyancy change, thereby forming a highly credible scientific decision-making basis and enabling a closed-loop feedback of prediction-interpretation-control. Attached Figure Description

[0039] The accompanying drawings, which are provided to further illustrate this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0040] Figure 1A flowchart of a shield tunnel segment buoyancy prediction method based on an improved hybrid neural network provided in this application. Detailed Implementation

[0041] Based on the background art, this application provides a shield tunnel segment buoyancy prediction system based on an improved hybrid neural network, comprising:

[0042] The data acquisition module collects construction parameters and structural response parameters in real time based on the tunnel boring machine data acquisition system and synchronous monitoring equipment.

[0043] The data processing module performs outlier removal, normalization, and principal component analysis on the construction parameters and structural response parameters to form input samples. Each parameter in the input samples is denoted as a feature input vector.

[0044] The tunnel boring machine (TBM) data acquisition system can record construction parameters and structural response parameters in real time. However, in this embodiment, the construction parameters determined by PCA include propulsion parameters, grouting parameters, attitude parameters, and historical buoyancy. In this embodiment, propulsion parameters include propulsion rate, total thrust, and cutterhead torque; grouting parameters include grouting pressure and grouting volume; attitude parameters include cutterhead penetration, soil chamber pressure, and screw conveyor speed; and structural response parameters include tail clearance and shield attitude angle.

[0045] use The principle is to remove outliers, specifically by calculating the mean of each type of data collected. and standard deviation It will fall within the interval [ , Data outside the defined range is considered outlier and removed. For data crossing stratigraphic fault zones and stratigraphic boundaries, if data collection is missing, interpolation is used to supplement the data based on data collected at adjacent times. In practice, the Min-Max standardization method is used to normalize the collected data.

[0046] After the above processing, Principal Component Analysis (PCA) is used to reduce the dimensionality of the high-dimensional input data. By setting a normalized sample matrix, calculating the covariance matrix, eigenvalue decomposition, and ranking, the top samples with a cumulative variance contribution rate exceeding 95% are retained. k In this embodiment, the principal component variables are typically taken as follows: k =A value between 10 and 12. Principal component variables can effectively characterize the core correlation features between shield tunneling state, grouting behavior and ground response, reducing the computational burden of the model.

[0047] In the GRU-SVM sub-model path, a time-series window structure needs to be constructed for the processed samples. In this embodiment, considering the significant time delay and inertial effect of the buoyancy response during shield tunneling, this application sets the time window length... T =48, meaning data from the past 4 hours, with each training sample consisting of continuous... T The input is composed of a sequence of feature data at each time step, i.e., the input is constructed as... , Indicates the first t Moment k Principal component variables, That is, the multivariate time series samples used as input for GRU neural network modeling.

[0048] The prediction module includes an XGBoost sub-model and a GRU-SVM sub-model.

[0049] The XGBoost sub-model takes the input sample as input and outputs the first boost prediction value. In the prediction module, the root mean square error of the XGBoost sub-model on the training set is used as the fitness function. An improved sparrow search algorithm is used to search for the hyperparameter combination that minimizes the root mean square error of the prediction as the optimal hyperparameter. The optimal XGBoost sub-model is constructed based on the optimal hyperparameter and used to output the first boost prediction value.

[0050] In practical modeling, let the input feature samples be... , representing the n construction characteristic variables (such as advance rate, grouting pressure, shield tail gap, etc.) of the i-th sample, with the corresponding measured value of the buoyancy. XGBoost uses multiple functions The regression tree represents the prediction of buoyancy, and its model form is as follows:

[0051]

[0052] in, This represents the predicted float value, where K is the number of regression trees. It is the function of the k-th regression tree.

[0053] The Improved Sparrow Search Algorithm (ISSA) simulates the "flocked foraging" behavior of sparrows in nature when optimizing the XGBoost model. The algorithm treats each candidate solution (i.e., a set of XGBoost hyperparameters, such as tree depth, learning rate, sampling ratio, etc.) as a "sparrow," and all sparrows form a "population" that continuously moves within the search space to find the location that minimizes the model's prediction error. The "discoverers" are the sparrows that lead the foraging search, responsible for exploring the entire search space and finding potential better solutions. The "followers" follow the discoverers, fine-tuning the hyperparameters in the local space and meticulously searching for possible optimal hyperparameters. The "guardians" act as the population's protectors; when the algorithm fails to find a better solution for an extended period, they randomly jump out of the current area to explore new, unknown regions, thus preventing the entire population from getting trapped in local optima.

[0054] In each iteration, each "sparrow" (i.e., a set of hyperparameters) is used to train an XGBoost model, and the prediction error is calculated on the validation set, measured by the root mean square error (RMSE). The algorithm compares the performance of all sparrows; those with smaller errors represent those that have found better food (i.e., better hyperparameters). As iterations proceed, the entire sparrow population gradually converges towards the region with the smallest error, eventually converging to a set of globally optimal hyperparameters.

[0055] In this way, ISSA can fully combine "global exploration" and "local fine-tuning", which can not only ensure that the hyperparameter combination with smaller error is found, but also avoid the problem that traditional algorithms are prone to getting trapped in local optima, thus significantly improving the accuracy and stability of the XGBoost model in shield tunneling buoyancy prediction tasks.

[0056] In the implementation of this invention, the hyperparameter combination of the XGBoost model is not merely a mathematical adjustment factor, but a core element that directly affects the buoyancy prediction model's ability to learn from and generalize engineering data. Specifically:

[0057] The maximum depth of the tree (max_depth) corresponds to the model's ability to characterize the complex interactions between construction variables. In tunnel boring machine (TBM) construction, factors such as advance rate, grouting pressure, and cutterhead torque often exhibit nonlinear superposition effects. A deeper tree can capture more complex relationships, but excessive depth may lead to overfitting to a particular working condition.

[0058] The learning rate (η) determines the magnitude of each model iteration update, analogous to the "sensitivity of parameter adjustment" in construction. If the learning rate is too large, like frequent and large fluctuations in grouting pressure, the prediction results will be extremely unstable; if it is too small, it is similar to overly sluggish adjustments in construction, resulting in a slow model convergence speed.

[0059] Subsampling rate (colsample_bytree): Controls how many samples and features the model uses in each tree construction. It's equivalent to "how to balance total information with random disturbances under limited monitoring data." In engineering scenarios, this is like how construction cannot rely on a single parameter but needs to extract representative features from different stages to make a comprehensive judgment.

[0060] Regularization coefficients (λ, α): These correspond to constraints on model complexity and can be seen as a manifestation of engineering safety redundancy. For example, when extreme values ​​occur in the grouting volume or shield tail gap, regularization can prevent the model from over-relying on outliers, thereby ensuring the robustness of the prediction results.

[0061] Minimum leaf weight (min_child_weight): This is the minimum amount of information that a decision tree must have when splitting. It is similar to the amount of data that must be accumulated in engineering monitoring in order to make a reliable judgment and avoid unreasonable prediction branches due to individual anomalies.

[0062] Number of trees (n_estimators): Similar to "multiple independent judgments and weighted results" in a construction site, the more trees there are, the more stable the model is, but too many trees will increase the amount of computation and the risk of overfitting.

[0063] Therefore, the set of hyperparameters obtained by ISSA optimization in this invention is essentially finding the best balance between "model learning ability" and "actual engineering laws": it can make full use of the complex relationship between multiple construction variables such as advance rate, grouting parameters, and shield tail gap, while avoiding overfitting to a single working condition, thereby ensuring the accuracy and robustness of the prediction model under different geological conditions.

[0064] The GRU-SVM sub-model takes input feature variables with temporal characteristics as input and outputs a second floating prediction value; the construction method of the GRU-SVM sub-model is as follows:

[0065] The state vector output by the GRU neural network is used as input to the support vector regression model for nonlinear fitting to output the second buoyancy prediction value.

[0066] Specifically, GRU (Gated Recurrent Unit) is mainly used to process the time-series characteristics of tunnel boring machine (TBM) construction data. Because the construction process is continuous, parameters such as the advance speed, grouting pressure, and earth pressure chamber pressure at each moment are affected by the previous moment, and changes in buoyancy often have a cumulative effect. Assuming at time... t Input feature vector x t x tIt includes variables such as advance rate, grouting pressure, and shield tail clearance. GRU processes x... t At that time, two "gating factors" will be calculated first:

[0067] Update Gate (zt): This determines how much past memory to retain and how much new information to introduce at this moment. In this embodiment, it can be compared to the process of tunnel boring machine (TBM) advancement, where workers need to determine whether "the current advancement speed and grouting situation are more important than the previous situation." If they are more important, more new information is introduced; if past experience is still valid, more historical information is retained.

[0068] Reset Gate (rt): Determines whether to "forget" some past states when combining current inputs. An engineering analogy is: if a tunnel boring machine enters a completely new geological stratum, the experience accumulated in soft soil may no longer be applicable. In this case, a partial "reset" is needed, relying more on new monitoring data.

[0069] Under the control of these two gates, the GRU will generate a candidate state. This is equivalent to a preliminary assessment of "how the buoyancy might change under current conditions." Next, the GRU will use an update gate... weighted fusion candidate states and past hidden states :

[0070] ;

[0071] in, This is the final output, which is GRU's understanding of the "buoyancy state under current construction conditions." It includes both the cumulative effects of history and the changes brought about by new inputs.

[0072] FSVR stands for Support Vector Regression Function Based on Kernel Mapping. In this application, FSVR refers to using kernel mapping technology to map the original input data from a "low-dimensional space" to a "high-dimensional space," and then using a smooth curve or hyperplane to fit the data's variation in the high-dimensional space. Specifically, suppose we can see the relationship between "propulsion rate" and "grouting pressure" and buoyancy in a two-dimensional plane, but this relationship may be very complex and difficult to fit directly with a suitable straight line in two-dimensional space. However, by mapping these two variables to a higher-dimensional space using a kernel function, the complex, curved relationship may become a near-straight line in this higher-dimensional space, at which point the optimal fitting function can be found using support vector machines.

[0073] The fusion module is used to fuse the first buoyancy prediction value and the second buoyancy prediction value into a buoyancy prediction value;

[0074] An interpretation output module calculates the marginal contribution value of each input feature variable in the input sample based on the predicted buoyancy value. The marginal contribution value is used to characterize the buoyancy of the tunnel segment under the influence of construction parameters.

[0075] Based on the above system, as shown in the appendix Figure 1 As shown, this application provides a method for predicting the buoyancy of tunnel segments based on an improved hybrid neural network, including the following steps:

[0076] S1. Collect construction parameters and structural response parameters;

[0077] S2. Perform outlier removal, normalization, and principal component analysis on the collected data;

[0078] In step S2, 3 Outlier removal is performed according to the following principles:

[0079] Calculate the mean of each type of data collection and standard deviation It will fall within the interval Data outside the range is treated as outliers and removed; for data crossing stratigraphic fault zones and stratigraphic boundaries, if the collected data is missing, interpolation is used to supplement the data based on data collected at adjacent times.

[0080] In step S2, the collected data is normalized using the Min-Max normalization method.

[0081] In step S2, the covariance matrix of the matrix formed by the normalized collected data is calculated and eigenvalue decomposition and sorting are performed. Principal component variables with a cumulative variance contribution rate greater than 95% are retained as comprehensive variables, and the comprehensive variables are used as input samples for the XGBoost sub-model. A time series window is added to the input samples of the XGBoost sub-model as input samples for the GRU-SVM sub-model.

[0082] S3. Based on the data collected in step S2, construct the training input samples of the XGBoost sub-model, input the training input samples into the XGBoost sub-model for training, use the root mean square error of prediction as the fitness function during the training process, use the improved sparrow search algorithm to search for the hyperparameter combination that minimizes the root mean square error of prediction as the optimal hyperparameter, construct the optimal XGBoost sub-model based on the optimal hyperparameter and use it to output the first boost prediction value.

[0083] S4. Based on the data collected in step S2, construct training input samples with temporal characteristics for the GRU-SVM sub-model. Input the training input samples into the output state vector of the GRU neural network, and input the state vector into the support vector regression model for training. Obtain the fitting function through kernel function mapping using nonlinear fitting. The fitting result is used as the second buoyancy prediction value. The object of nonlinear fitting is the nonlinear relationship between the state vector output by GRU (input features) and the segment buoyancy (target value). The state vector output by the GRU neural network after processing the temporal data is essentially the result of extracting the temporal features of the original shield construction parameters (such as cutterhead torque, grouting pressure, etc.) (including the time dimension dependency). However, the relationship between these state vectors and the segment buoyancy is often not a simple linear relationship. For example, a small change in grouting pressure may cause a sharp increase in buoyancy in one stage, while having a weak effect in another stage. Therefore, it is necessary to use the kernel function of SVM to perform nonlinear mapping to fit this complex relationship, and finally output the prediction value based on this relationship, which is the second buoyancy prediction value.

[0084] S5. The first and second floating prediction values ​​are fused using the inverse error weighting method to obtain the floating prediction value. In step S5, the root mean square error of the predictions of the XGBoost sub-model and the GRU-SVM sub-model on the validation set is first calculated. and Then calculate the predicted buoyancy value according to the following formula:

[0085] ;

[0086] In the formula, This represents the predicted rise value. This represents the first predicted rise value. This represents the second predicted rise value.

[0087] S6. The SHAP interpretation algorithm is used to construct a linear weighted model of the input samples and their corresponding buoyancy prediction values. A baseline prediction value is set in the linear weighted model, and the marginal contribution value corresponding to each construction parameter in the input samples is calculated.

[0088] In step S6, the linear weighted model is:

[0089] ;

[0090] In the formula, Indicates the baseline forecast value. Indicates the first i The marginal contribution of each construction parameter to the predicted buoyancy value.

[0091] For any parts not mentioned in this application, existing technologies may be used or referenced.

[0092] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of this application should be included within the scope of the claims of this application.

Claims

1. A shield segment floating lifting prediction system based on an improved hybrid neural network, characterized in that, The system comprises: a data acquisition module that acquires construction parameters and structural response parameters in real time based on a shield tunneling machine data acquisition system and a synchronous monitoring device; a data processing module that performs outlier rejection, normalization and principal component analysis operations on the construction parameters and structural response parameters to form input samples, each parameter in the input samples being recorded as a feature input vector; a prediction module that comprises an XGBoost submodel and a GRU-SVM submodel; the XGBoost submodel takes the input samples as input and outputs a first floating and lifting prediction value; the GRU-SVM submodel takes a feature input vector with time sequence characteristics as input and outputs a second floating and lifting prediction value; in the prediction module, the root mean square error of the XGBoost submodel on the training set is used as a fitness function, and an improved sparrow search algorithm is used to search for a combination of hyperparameters that minimizes the root mean square error, so as to obtain optimal hyperparameters; based on the optimal hyperparameters, an optimal XGBoost submodel is constructed and used to output the first floating and lifting prediction value; in the prediction module, the GRU-SVM submodel is constructed in the following manner: the state vector output by the GRU neural network is input into a support vector regression model for nonlinear fitting to output the second floating and lifting prediction value; a fusion module that is configured to fuse the first floating and lifting prediction value and the second floating and lifting prediction value into a floating and lifting prediction value; an explanation output module that calculates the marginal contribution value of each feature input vector in the input sample based on the floating and lifting prediction value, the marginal contribution value being used to represent the segment floating and lifting amount under the influence of the construction parameters.

2. A shield segment floating lifting prediction method based on an improved hybrid neural network, characterized in that, The method is implemented based on the system of claim 1 and comprises the following steps: S1, acquiring construction parameters and structural response parameters; S2, performing outlier rejection, normalization and principal component analysis operations on the acquired data to establish an input sample comprising feature input vectors; S3, constructing a training input sample for the XGBoost submodel based on the acquired data of step S2, inputting the training input sample into the XGBoost submodel for training, using the root mean square error of prediction as a fitness function during the training process, using an improved sparrow search algorithm to search for a combination of hyperparameters that minimizes the root mean square error, obtaining optimal hyperparameters based on the optimal hyperparameters, constructing an optimal XGBoost submodel based on the optimal hyperparameters and using the optimal XGBoost submodel to output a first floating and lifting prediction value; S4, constructing a training input sample with time sequence characteristics for the GRU-SVM submodel based on the acquired data of step S2, inputting the training input sample into a GRU neural network to output a state vector, inputting the state vector into a support vector regression model for training, performing nonlinear fitting through kernel function mapping to obtain a fitting function, and using the fitting result as a second floating and lifting prediction value; S5, fusing the first floating and lifting prediction value and the second floating and lifting prediction value using an error reciprocal weighting method to obtain a floating and lifting prediction value; S6, using a SHAP explanation algorithm to construct a linear weighted model from the input sample and its corresponding floating and lifting prediction value, setting a baseline prediction value in the linear weighted model, and calculating the marginal contribution value of each construction parameter in the input sample.

3. The shield segment floating lifting prediction method based on the improved hybrid neural network according to claim 2, characterized in that, In step S2, the 3 Principle of outlier rejection is adopted, specifically: The mean value of each acquisition data is calculated and the standard deviation Data falling outside the interval is regarded as an abnormal value and is removed; for data crossing the stratigraphic fault zone or the formation boundary, if the acquisition data is missing, the data is supplemented by interpolation based on the adjacent time acquisition data; In step S2, the collected data is normalized by using the Min-Max normalization method.

4. The shield segment floating lifting prediction method based on the improved hybrid neural network according to claim 3, characterized in that, In step S2, the covariance matrix of the matrix formed by the normalized collected data is calculated and subjected to eigenvalue decomposition and sorting, and the principal component variables with the cumulative variance contribution rate greater than 95% are retained as the comprehensive variables, and the comprehensive variables are taken as the input samples of the XGBoost sub-model; The time sequence window is added in the input samples of the XGBoost sub-model to serve as the input samples of the GRU-SVM sub-model.

5. The shield segment floating lifting prediction method based on the improved hybrid neural network according to claim 4, characterized in that, In step S5, first, the root mean square error values of the prediction of the XGBoost sub-model and the GRU-SVM sub-model on the validation set are calculated and and then the buoyancy prediction value is calculated according to the following formula: ； wherein denotes the lift forecast value, denotes the first lift forecast value, denotes the second lift forecast value.

6. The shield segment floating lifting prediction method based on the improved hybrid neural network according to claim 5, characterized in that, In step S6, the linear weighting model is: ； wherein represents the reference prediction value, represents the marginal contribution value of the i construction parameter to the floating prediction value.

Images