Shield attitude prediction method based on pseudo residual hybrid model
By combining a pseudo-residual hybrid model with the dynamic fusion of multi-head self-attention and multiple regression trees, the problem of insufficient accuracy in shield tunnel attitude prediction under complex geological conditions is solved, and high-precision, real-time shield tunnel attitude prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-07
AI Technical Summary
Existing shield tunnel attitude prediction methods lack dynamic adaptive mechanisms, making it difficult to quickly optimize when monitoring environmental changes in real time. This results in difficulty in achieving rapid response and environmental adaptation to sudden changes, especially under complex geological conditions where prediction accuracy is insufficient.
A pseudo-residual hybrid model is adopted. By constructing a piecewise correction shield attitude prediction model driven by global trend pseudo-residual, and combining a multi-head self-attention mechanism and a multi-regression tree model, the global trend and local details are dynamically integrated to achieve high-precision prediction of shield attitude.
It achieves high-precision shield tunnel attitude prediction under complex geological conditions, with lower cost, higher generalization ability and computational efficiency, and can sensitively capture minute fluctuations to achieve real-time prediction.
Smart Images
Figure CN121808358A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of tunnel boring machine (TBM) construction technology, and in particular to a method for predicting TBM attitude based on a pseudo-residual hybrid model. Background Technology
[0002] During tunnel boring machine (TBM) construction, excessive deviation in the TBM's attitude can easily lead to engineering risks such as tunnel alignment deviation, structural collisions, or ground settlement. Therefore, predicting the TBM's attitude is crucial. Previous TBM attitude prediction methods were mostly based on machine learning algorithms, which provide more accurate prediction results compared to traditional methods such as empirical and numerical methods. However, existing methods still have some shortcomings: they lack dynamic adaptive mechanisms, making it difficult to quickly optimize when monitoring environmental changes in real time, thus hindering rapid response and environmental adaptation to sudden changes.
[0003] Therefore, there is an urgent need to provide a pseudo-residual multi-model dynamic fusion correction mechanism that not only focuses on the model's ability to capture future trends in the data, but also emphasizes the sensitivity to small fluctuations, so as to achieve high-precision prediction of the future shield tunnel attitude under complex geological conditions. Summary of the Invention
[0004] This invention provides a shield tunnel attitude prediction method based on a pseudo residual hybrid model to overcome the above-mentioned technical problems.
[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A shield tunneling attitude prediction method based on a pseudo-residual hybrid model includes the following steps: S1: Collect and acquire historical data of the tunnel boring machine (TBM) during the tunneling process; the historical data includes time-series data of TBM tunneling control parameters, TBM attitude parameters, and TBM tunneling geological parameters; S2: Conduct correlation analysis on the shield tunneling control parameters and shield attitude parameters, and select parameter features that meet the preset correlation threshold to establish a dataset; S3: Perform data preprocessing on the dataset and obtain the preprocessed data; S4: Building based on Global trend pseudo-residual driven A pseudo-residual hybrid prediction model for segmented shield tunneling attitude correction includes an input module and a module based on... The model's first pose prediction module, based on The model includes a second attitude prediction module, a dynamic fusion correction module based on global trend pseudo residuals, and an output module; The input module is used to transmit the preprocessed data and the corresponding shield tunneling geological parameters to the first attitude prediction module and the second attitude prediction module, respectively. The first attitude prediction module is used to perform patch segmentation and feature mapping on the output of the input module to obtain several feature space vectors, and extract the global temporal dependency between different feature space vectors based on the multi-head self-attention mechanism to obtain several global dependency feature vectors. At the same time, the global dependency feature vectors are fused to obtain global trend features, and the shield attitude at future moments is predicted based on the global trend features to obtain the first prediction result. The second attitude prediction module is used to determine the attitude based on the output of the input module. The model extracts nonlinear time-series relationships to obtain local trend features, and predicts the shield attitude at future moments based on the local trend features to obtain a second prediction result. The dynamic fusion correction module is used to define a global trend pseudo residual based on the first prediction result and the second prediction result, and to construct a segmented mapping strategy based on the global trend pseudo residual to obtain the shield attitude correction prediction result at future time and output it through the output module. S5: The pseudo-residual hybrid prediction model is trained based on the preprocessed data and the geological parameters of the tunnel boring machine to obtain the optimal prediction model and achieve the prediction of the tunnel boring machine attitude.
[0006] Furthermore, the shield tunneling control parameters mentioned in S1 include shield tail seal pressure data, hydraulic cylinder thrust data, and face excavation resistance data; the shield tail seal pressure data includes at least the sensor timing data corresponding to the front, middle, and rear of the shield tail seal and the grouting pressure; the hydraulic cylinder thrust data includes at least the timing data of the propulsion AF group pressure, propulsion AF group displacement, propulsion pressure, total propulsion force, and propulsion speed; the face excavation resistance data includes at least the timing data of the soil chamber pressure, cutterhead rotation speed, cutterhead torque, and penetration depth. The shield attitude parameters include time-series data of the horizontal deviation between the shield head and the shield tail, the vertical deviation between the shield head and the shield tail, and the roll angle radian of the shield, which correspond to the shield tunneling control parameters. The geological parameters for shield tunneling include at least the percentage content of silty clay, slightly weathered fractured granite, moderately weathered tuff, strongly weathered tuff, moderately weathered fractured limestone, breccia, and moderately weathered fine rock.
[0007] Furthermore, the expression for the correlation analysis between the shield tunneling control parameters and the shield attitude parameters in S2 is as follows:
[0008] In the formula: This represents the actual values of the tunnel boring machine (TBM) parameters; This represents the actual value of the shield tunneling attitude parameters; This represents the total number of time-series data corresponding to the shield tunneling control parameters and shield attitude parameters; This indicates the correlation value.
[0009] Furthermore, the data preprocessing method described in S3 includes the following steps: S31: Remove tunnel boring machine shutdown data from the dataset based on empirical values to obtain optimized data; S32: Remove outliers from the optimized data and obtain intermediate data using the Local Outlier Factor (LOF) algorithm; S33: Based on the Discrete Wavelet Transform (DWT) method, the intermediate data is denoised to obtain preprocessed data.
[0010] Furthermore, in S4, based on The method for obtaining the first prediction result in the model's first pose prediction module is as follows: S41: Based on a preset time series window of length L, divide the preprocessed data and the time series data of the corresponding shield tunneling geological parameters to obtain several time series data segments; S42: Convert the time series windows corresponding to each time series data segment. Divided into A patch block of equal length and , Where P represents the position and length of the patch block; Indicates the total number of patch blocks; Represents the space of real numbers; Indicates the first One patch block; S43: By linear transformation matrix With position encoding Map each patch block to the feature space to obtain the feature space vector; And the formula for obtaining the feature space vector is:
[0011] In the formula: Indicates the first The mapping features corresponding to each patch block; Describes the linear transformation matrix and ; Represents the dimension of the feature space; Indicates the first The location information of each patch block and ; Indicates the corresponding time series window eigenspace vectors; S44: Define and obtain the input feature matrix based on the feature space vectors of each time series data segment. and ; S45: Extract the global temporal dependency between different feature space vectors based on the multi-head self-attention mechanism, and obtain several global dependency feature vectors; And the expression for the multi-head self-attention mechanism is:
[0012] In the formula: Represents the learnable projection matrix; Represents the query vector; Key vector; Represents a value vector; Indicates transpose; Indicates feature dimension; S46: Fuse the global dependent feature vectors, that is, perform element-wise summation and obtain global trend features, predict the shield attitude at future moments based on global trend features, and obtain the first prediction result; And the formula for obtaining the first prediction result is:
[0013] In the formula: Indicates the prediction step size; This represents the set of weight parameters for the first attitude prediction module; Indicates global trend characteristics; This indicates the predicted attitude of the tunnel boring machine at a future moment.
[0014] Furthermore, the formula for obtaining the second prediction result in S4 is:
[0015] In the formula: express The first in the model A regression tree; This indicates the corresponding tree structure and leaf node weights; express Preprocessed data and geological parameters of shield tunneling at each moment.
[0016] Furthermore, the method for obtaining the shield attitude correction prediction results for future moments in S4 is as follows: Using the first prediction result as the baseline, the global trend pseudo residual is defined based on the second prediction result; And the expression for the global trend pseudo residual is:
[0017] In the formula: This indicates the first prediction result; This indicates the second prediction result; This indicates a pseudo-residual representing the global trend. The piecewise mapping strategy is constructed based on the global trend pseudo residual, and its expression is:
[0018] In the formula: This indicates the predicted attitude correction results for the tunnel boring machine at future moments; , Indicates the adaptive threshold parameter; , This represents the learnable correction coefficient; The shield attitude correction prediction results for future moments are obtained based on the segmented mapping strategy and output through the output module.
[0019] Furthermore, the method for obtaining the optimal prediction model in S5 is as follows: S51: Use the preprocessed data and shield tunneling geological parameters as sample data, and divide the training set and validation set according to a preset ratio; S52: Train the pseudo-residual mixture prediction model based on the training set to obtain the trained pseudo-residual mixture prediction model; S53: Based on the model loss function, the trained pseudo-residual mixture prediction model is validated using a validation set; That is, to determine whether the output of the pseudo-residual mixture prediction model after training converges; If the output of the trained pseudo-residual fusion prediction model converges, then the trained pseudo-residual fusion prediction model is confirmed to be the optimal prediction model. Otherwise, based on the Bayesian optimization algorithm, the weight parameters of the trained pseudo-residual mixture prediction model are adaptively adjusted, and step S52 is repeated until the weight parameters of the trained pseudo-residual mixture prediction model with converged output are confirmed to be the optimal weight parameters, and the pseudo-residual mixture prediction model is reconstructed to obtain the optimal prediction model.
[0020] Furthermore, the expression for the model loss function described in S53 is:
[0021]
[0022] In the formula: Represents the model loss function; The loss function of the first attitude prediction module is represented, which includes, but is not limited to, the mean squared error function. This represents the loss function output by the dynamic fusion correction module based on global trend pseudo residuals, which includes, but is not limited to, the mean squared error function. This represents the loss function of the second pose prediction module; Represents the loss function term; Represents the regularization term; Indicates the number of controllable leaf nodes; This represents a function term used to control the magnitude of the output value of the leaf nodes; This represents the output value of the j-th leaf node; This represents the weighting coefficient.
[0023] Beneficial effects: This invention provides a shield tunneling attitude prediction method based on a pseudo-residual hybrid model, by constructing a method based on... Global trend pseudo-residual driven A pseudo-residual hybrid prediction model for segmented correction of shield attitude, combined with Model and This invention leverages the advantages of the model and introduces a global trend pseudo-residual, dynamically fusing the two to predict the future attitude of the tunnel boring machine (TBM), thus achieving real-time prediction. It overcomes the limitations of a single theory, possessing advantages such as lower cost, greater generalization ability, higher computational efficiency, and superior prediction accuracy. By capturing future trends in data and emphasizing sensitivity to minute fluctuations, it achieves high-precision prediction of the future attitude of the TBM under complex geological conditions. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 The flowchart shows the shield tunneling attitude prediction method based on the pseudo-residual hybrid model of the present invention. Figure 2 This is a block diagram illustrating the core technical roadmap of the method described in this embodiment; Figure 3 This is a comparison curve chart showing the removal of tunnel boring machine shutdown data in this embodiment; Figure 4 This is a comparison curve chart showing the removal of abnormal data in this embodiment; Figure 5 This is a comparison curve of data denoising in this embodiment; Figure 6 For this embodiment A schematic diagram of the model's input and output settings; Figure 7 For this embodiment A schematic diagram of the model's input and output settings; Figure 8A flowchart illustrating the input and output settings in this embodiment; Figure 9 This is a comparison chart of the actual values and the prediction results of the method described in this embodiment; Figure 10 The method described in this embodiment, Model and Comparison chart of model prediction results; Figure 11 This is a comparison chart of evaluation metrics for different models in this embodiment. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] This embodiment provides a shield tunneling attitude prediction method based on a pseudo-residual hybrid model, such as Figure 1 To TuTu Figure 2 As shown, the specific steps include: S1: Collect and acquire historical data of the tunnel boring machine (TBM) during the tunneling process; the historical data includes time-series data of TBM tunneling control parameters, TBM attitude parameters, and TBM tunneling geological parameters; Specifically, the shield tunneling control parameters include tail seal pressure data, hydraulic cylinder thrust data, and face excavation resistance data; the tail seal pressure data includes at least the time-series data of sensors corresponding to the front, middle, and rear of the tail seal and the grouting pressure; the hydraulic cylinder thrust data includes at least the time-series data of the propulsion AF group pressure, propulsion AF group displacement, propulsion pressure, total propulsion force, and propulsion speed; the face excavation resistance data includes at least the time-series data of the soil chamber pressure, cutterhead rotation speed, cutterhead torque, and penetration depth. The shield attitude parameters include time-series data of the horizontal deviation between the shield head and the shield tail, the vertical deviation between the shield head and the shield tail, and the roll angle radian of the shield, which correspond to the shield tunneling control parameters. The geological parameters for shield tunneling include at least the percentage content of silty clay, slightly weathered fractured granite, moderately weathered tuff, strongly weathered tuff, moderately weathered fractured limestone, breccia, and moderately weathered fine rock. S2: Conduct correlation analysis on the shield tunneling control parameters and shield attitude parameters, and select parameter features that meet the preset correlation threshold to establish a dataset; Specifically, the Pearson correlation coefficient is used to identify the correlation between tunneling parameters and target shield attitude parameters, and its expression is as follows:
[0028] In the formula: This represents the actual values of the tunnel boring machine (TBM) parameters; This represents the actual value of the shield tunneling attitude parameters; This represents the total number of time-series data corresponding to the shield tunneling control parameters and shield attitude parameters; Indicates the correlation value; S3: Perform data preprocessing on the dataset and obtain the preprocessed data; Specifically, the data preprocessing method is as follows: S31: Remove tunnel boring machine shutdown data from the dataset based on empirical values to obtain optimized data; like Figure 3 As shown, for example, the cutter head rotation speed: Figure 3 (a) indicates that data prior to shutdown has been removed; Figure 3 (b) indicates the removal of data after shutdown; S32: Remove outliers from the optimized data and obtain intermediate data using the Local Outlier Factor (LOF) algorithm; like Figure 4 As shown, for example, the cutter head rotation speed: Figure 4 (c) represents the data before outliers were removed; Figure 4 (d) represents the data after removing outliers; S33: Based on the Discrete Wavelet Transform (DWT) method, the intermediate data is denoised to obtain preprocessed data, such as... Figure 5 As shown in the figure. In this embodiment, during the tunnel boring machine (TBM) excavation process, the time-series data contains a large number of TBM shutdown periods. Therefore, it is necessary to first remove shutdown data to improve the accuracy of attitude research. Subsequently, outliers caused by sensor malfunctions are identified and removed using the Local Outlier Factor (LOF). Simultaneously, the data is often affected by environmental noise and measurement errors; therefore, Discrete Wavelet Transform (DWT) is applied to denoise the original data to improve the stability and reliability of subsequent analysis. After completing the data preprocessing, interfaces are built for the two models according to the input and output requirements, and settings are defined for each model. and Input and output such as Figures 6 to 8 As shown; S4: Building based on Global trend pseudo-residual driven A pseudo-residual hybrid prediction model for segmented shield tunneling attitude correction includes an input module and a module based on... The model's first pose prediction module, based on The model includes a second attitude prediction module, a dynamic fusion correction module based on global trend pseudo residuals, and an output module; The input module is used to transmit the preprocessed data and the corresponding shield tunneling geological parameters to the first attitude prediction module and the second attitude prediction module, respectively. The first attitude prediction module is used to perform patch segmentation and feature mapping on the output of the input module to obtain several feature space vectors, and extract the global temporal dependency between different feature space vectors based on the multi-head self-attention mechanism to obtain several global dependency feature vectors. At the same time, the global dependency feature vectors are fused to obtain global trend features, and the shield attitude at future moments is predicted based on the global trend features to obtain the first prediction result. Specifically, the method for obtaining the first prediction result is as follows: S41: Based on a preset time series window of length L, divide the preprocessed data and the time series data of the corresponding shield tunneling geological parameters to obtain several time series data segments; S42: To efficiently process long-term series data, the time series windows corresponding to each time series data segment are... Divided into A patch block of equal length and , Where P represents the position and length of the patch block; Indicates the total number of patch blocks; Represents the space of real numbers; Indicates the first One patch block; S43: By linear transformation matrix With position encoding Map each patch block to the feature space to obtain the feature space vector; And the formula for obtaining the feature space vector is:
[0029] In the formula: Indicates the first The mapping features corresponding to each patch block; Describes the linear transformation matrix and ; Represents the dimension of the feature space; Indicates the first The location information of each patch block and Position encoding enables the model to perceive the order between patches; Indicates the corresponding time series window eigenspace vectors; S44: Define and obtain the input feature matrix based on the feature space vectors of each time series data segment. and ; S45: Extract the global temporal dependency between different feature space vectors based on the multi-head self-attention mechanism, and obtain several global dependency feature vectors; And the expression for the multi-head self-attention mechanism is:
[0030] In the formula: Represents the learnable projection matrix; Represents the query vector; Key vector; Represents a value vector; Indicates transpose; Indicates feature dimension; S46: Fuse the global dependent feature vectors, that is, perform element-wise summation and obtain global trend features, predict the shield attitude at future moments based on global trend features, and obtain the first prediction result; And the formula for obtaining the first prediction result is:
[0031] In the formula: Indicates the prediction step size; This represents the set of weight parameters for the first attitude prediction module; Indicates global trend characteristics; This indicates the predicted attitude of the tunnel boring machine at a future moment.
[0032] In this embodiment, the real-time shield tunnel attitude prediction model, namely the first attitude prediction module, utilizes... The model divides the input time series of length L into multiple equal-length patch blocks, performs parallel positional encoding, and then combines a multi-head self-attention mechanism to mine the long-distance dependencies between patches, capturing long-term temporal dependencies and global trend features.
[0033] The second attitude prediction module is used to determine the attitude based on the output of the input module. The model extracts nonlinear temporal relationships from time-series data to obtain local trend features, and predicts the shield attitude at future moments based on these local trend features to obtain a second prediction result. Specifically, the formula for obtaining the second prediction result is:
[0034] In the formula: express The first in the model A regression tree; This indicates the corresponding tree structure and leaf node weights; express The preprocessed data and shield tunneling geological parameters corresponding to the given time; the shield attitude prediction model in this embodiment, namely the second attitude prediction module, utilizes... The model captures nonlinear relationships and localized sharp fluctuations in time series data by iteratively fitting the residuals and gradually accumulating the predictions of multiple regression trees.
[0035] The dynamic fusion correction module is used to define a global trend pseudo residual based on the first prediction result and the second prediction result, and to construct a segmented mapping strategy based on the global trend pseudo residual to obtain the shield attitude correction prediction result at future time and output it through the output module. Specifically, this embodiment introduces a deep trend model. The prediction result, i.e., the first prediction result, is taken as the baseline, and is defined as follows: The global trend pseudo residual of the model's prediction result, i.e., the second prediction result, is:
[0036] In the formula: This indicates the first prediction result; This indicates the second prediction result; This indicates a pseudo-residual representing the global trend. Furthermore, a piecewise mapping strategy is constructed by dynamically adjusting the contribution weights of the two models based on the magnitude of the global trend pseudo-residual, aiming at dynamic fusion. The model's global trend prediction capability and The model's ability to correct local details is expressed by the segmented mapping strategy as follows:
[0037] In the formula: This indicates the predicted attitude correction results for the tunnel boring machine at future moments; , Indicates the adaptive threshold parameter; , This represents the learnable correction coefficients; the shield attitude correction prediction results for future moments are obtained according to the piecewise mapping strategy and output through the output module: when the pseudo-residual amplitude is small, The overall trend is more accurate, and a weighted average of the two should be used to maintain stability; when the spurious residual amplitude is large, It has an advantage in capturing extreme values / outliers, and its correction weights should be increased to better describe local details.
[0038] S5: Based on the preprocessed data and the geological parameters of the tunnel boring machine, the pseudo-residual hybrid prediction model is trained to obtain the optimal prediction model for predicting the tunnel boring machine's attitude. Specifically: S51: Use the preprocessed data and shield tunneling geological parameters as sample data, and divide the training set and validation set according to a preset ratio; S52: Train the pseudo-residual mixture prediction model based on the training set to obtain the trained pseudo-residual mixture prediction model; S53: Based on the model loss function, the trained pseudo-residual mixture prediction model is validated using a validation set. The expression for the model loss function is:
[0039]
[0040] In the formula: Represents the model loss function; The loss function of the first attitude prediction module is represented, which includes, but is not limited to, the mean squared error function. This represents the loss function output by the dynamic fusion correction module based on global trend pseudo residuals, which includes, but is not limited to, the mean squared error function. This represents the loss function of the second pose prediction module; Represents the loss function term; This represents the regularization term, used to avoid overfitting of the model. A regularization term was added when constructing the loss function; Indicates the number of controllable leaf nodes; This represents a function term used to control the magnitude of the output value of the leaf nodes; This represents the output value of the j-th leaf node; Indicates the weighting coefficient; That is, to determine whether the output of the pseudo-residual mixture prediction model after training converges; If the output of the trained pseudo-residual fusion prediction model converges, then the trained pseudo-residual fusion prediction model is confirmed to be the optimal prediction model. Otherwise, based on the Bayesian optimization algorithm ( The weight parameters of the trained pseudo-residual fusion prediction model are adaptively adjusted, and step S52 is repeated until the weight parameters of the convergent trained pseudo-residual fusion prediction model are confirmed to be optimal. The pseudo-residual fusion prediction model is then reconstructed to obtain the optimal prediction model. This embodiment adjusts the globally fused parameters in the model... Model parameters and The core hyperparameters of the model are set within a reasonable optimization range, and the Bayesian algorithm is used to search for the optimal hyperparameters, so that the prediction results during the training phase are closer to the actual results.
[0041] In this embodiment, the prediction results of the multi-model fusion correction mechanism based on pseudo-residues are compared and analyzed with the results of two single models, and the performance of the three models is evaluated using the following indicators: coefficient of determination (R²). 2 The mean absolute percentage error (MAPE), mean square error (MSE), mean absolute error (MAE), and bias (MBE) are all calculated using these methods.
[0042] like Figure 9 The figure shows a comparison and analysis of the real-time predicted values and actual values predicted by the method described in this embodiment. Square symbols represent the actual values, and circular symbols represent the future predicted values. The first 200 minutes represent the historical attitude values of the predicted target, and the latter 200 minutes are compared with the actual observed values within the same time interval. This shows that the overlap between the actual values and the future predicted values is significantly high, indicating that the model not only has the ability to predict the future attitude data of the tunnel boring machine in real time, but also exhibits strong robustness in terms of prediction accuracy. During training, a Bayesian optimization algorithm is used to find the optimal hyperparameters, including globally fused relevant parameters. Model parameters and Model parameters. Specific values for this embodiment are shown in Tables 1 to 3: Table 1. Optimal Hyperparameter Settings for Global Fusion
[0043] Table 2. Optimal hyperparameter settings for the model
[0044] Table 3. Optimal hyperparameter settings for the model
[0045] The method described in this embodiment is compared with that used in this embodiment. and The prediction results of the two single models are compared and analyzed. The comparison results in this embodiment are as follows: Figure 10 As shown, the straight line represents the true value, and the dashed line represents the true value. Predicted values and point lines are The predicted values and dashed lines are corrected predictions based on dynamic fusion of pseudo-residual multi-models. Therefore, we can conclude that: The model exhibits relatively stable residuals across most parameters, effectively capturing the general trends of the parameters. The model exhibits significant residual fluctuations but excels at capturing local parameter features. The pseudo-residual multi-model dynamic fusion correction model shows the smallest overall residual and a more stable distribution, demonstrating a significant advantage in capturing local details and eliminating systematic errors. Simultaneously, the shield tunneling attitude prediction model based on pseudo-residual dynamic fusion correction, i.e., the optimal prediction model, is compared with a control model. and Two single-model prediction results were systematically evaluated; and the real-time shield tunneling attitude prediction model with the highest performance level was selected based on the evaluation results. The predictive performance of the model was evaluated using the coefficient of determination, mean absolute percentage error, mean square error, mean absolute error, and bias. Figure 11 The comparison shown illustrates the performance of the pseudo-residual multi-model dynamic fusion correction model and the two-single model across various metrics. The squares represent... Curves, circles The curves, specifically the triangle, represent the pseudo-residual fusion model. This indicates that the pseudo-residual fusion model outperforms the two-single model in most metrics, but at the roll angle... The error is lowest because the roll angle range tends to be constant, reducing the need for complex modeling; It is sensitive to small fluctuations and prone to prediction errors under constant background conditions. This indicator shows that... Can inhibit Despite overfitting, the pseudo-residual multi-model dynamic fusion model still outperforms the others on this task. In summary, the method described in this embodiment can realistically simulate the prediction of shield tunneling attitude in actual construction environments and achieves real-time, high-precision prediction of shield tunneling attitude under complex geological conditions. Based on the experimental equipment involved in this invention, a method based on... Global trend pseudo-residual driven The dynamic fusion shield attitude prediction method provides a new approach for predicting the future attitude of shields in complex geological conditions.
[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A shield tunneling attitude prediction method based on a pseudo-residual hybrid model, characterized in that, Specifically, the following steps are included: S1: Collect and acquire historical data of the tunnel boring machine (TBM) during the tunneling process; the historical data includes time-series data of TBM tunneling control parameters, TBM attitude parameters, and TBM tunneling geological parameters; S2: Conduct correlation analysis on the shield tunneling control parameters and shield attitude parameters, and select parameter features that meet the preset correlation threshold to establish a dataset; S3: Perform data preprocessing on the dataset and obtain the preprocessed data; S4: Building based on Global trend pseudo-residual driven A pseudo-residual hybrid prediction model for segmented shield tunneling attitude correction includes an input module and a module based on... The model's first pose prediction module, based on The model includes a second attitude prediction module, a dynamic fusion correction module based on global trend pseudo residuals, and an output module; The input module is used to transmit the preprocessed data and the corresponding shield tunneling geological parameters to the first attitude prediction module and the second attitude prediction module, respectively. The first attitude prediction module is used to perform patch segmentation and feature mapping on the output of the input module to obtain several feature space vectors, and extract the global temporal dependency relationship between different feature space vectors based on the multi-head self-attention mechanism to obtain several global dependency feature vectors. At the same time, the global dependency feature vectors are fused to obtain global trend features, and the shield attitude at future moments is predicted based on the global trend features to obtain the first prediction result. The second attitude prediction module is used to determine the attitude based on the output of the input module. The model extracts nonlinear time-series relationships to obtain local trend features, and predicts the shield attitude at future moments based on the local trend features to obtain a second prediction result. The dynamic fusion correction module is used to define a global trend pseudo residual based on the first prediction result and the second prediction result, and to construct a segmented mapping strategy based on the global trend pseudo residual to obtain the shield attitude correction prediction result at future time and output it through the output module. S5: The pseudo-residual hybrid prediction model is trained based on the preprocessed data and the geological parameters of the tunnel boring machine to obtain the optimal prediction model and achieve the prediction of the tunnel boring machine attitude.
2. The shield tunneling attitude prediction method based on a pseudo-residual hybrid model according to claim 1, characterized in that, The shield tunneling control parameters described in S1 include tail seal pressure data, hydraulic cylinder thrust data, and face excavation resistance data. The tail seal pressure data includes at least the time-series data of sensors corresponding to the front, middle, and rear of the tail seal and the grouting pressure. The hydraulic cylinder thrust data includes at least the time-series data of the propulsion AF group pressure, propulsion AF group displacement, propulsion pressure, total propulsion force, and propulsion speed. The face excavation resistance data includes at least the time-series data of the soil chamber pressure, cutterhead rotation speed, cutterhead torque, and penetration depth. The shield attitude parameters include time-series data of the horizontal deviation between the shield head and the shield tail, the vertical deviation between the shield head and the shield tail, and the roll angle radian of the shield, which correspond to the shield tunneling control parameters. The geological parameters for shield tunneling include at least the percentage content of silty clay, slightly weathered fractured granite, moderately weathered tuff, strongly weathered tuff, moderately weathered fractured limestone, breccia, and moderately weathered fine rock.
3. The shield tunneling attitude prediction method based on a pseudo-residual hybrid model according to claim 2, characterized in that, The expression for the correlation analysis between shield tunneling control parameters and shield attitude parameters in S2 is as follows: In the formula: This represents the actual values of the tunnel boring machine (TBM) excavation parameters; This represents the actual value of the shield tunneling attitude parameters; This represents the total number of time-series data corresponding to the shield tunneling control parameters and shield attitude parameters; This indicates the correlation value.
4. The shield tunneling attitude prediction method based on a pseudo-residual hybrid model according to claim 3, characterized in that, The data preprocessing method described in S3 includes the following steps: S31: Remove tunnel boring machine shutdown data from the dataset based on empirical values to obtain optimized data; S32: Remove outliers from the optimized data and obtain intermediate data using the Local Outlier Factor (LOF) algorithm; S33: Based on the Discrete Wavelet Transform (DWT) method, the intermediate data is denoised to obtain preprocessed data.
5. The shield tunneling attitude prediction method based on a pseudo-residual hybrid model according to claim 4, characterized in that, S4 is based on The method for obtaining the first prediction result in the model's first pose prediction module is as follows: S41: Based on a preset time series window of length L, divide the preprocessed data and the time series data of the corresponding shield tunneling geological parameters into several time series data segments; S42: Convert the time series windows corresponding to each time series data segment. Divided into A patch block of equal length and , Where P represents the position and length of the patch block; Indicates the total number of patch blocks; Represent the space of real numbers; Indicates the first One patch block; S43: By linear transformation matrix With position encoding Map each patch block to the feature space to obtain the feature space vector; And the formula for obtaining the feature space vector is: In the formula: Indicates the first The mapping features corresponding to each patch block; Describes the linear transformation matrix and ; Represents the dimension of the feature space; Indicates the first The location information of each patch block and ; Indicates the corresponding time series window eigenspace vectors; S44: Define and obtain the input feature matrix based on the feature space vectors of each time series data segment. and ; S45: Extracting the input feature matrix based on a multi-head self-attention mechanism The global temporal dependencies between different feature space vectors within the region are used to obtain several globally dependent feature vectors. And the expression for the multi-head self-attention mechanism is: In the formula: Represents the learnable projection matrix; Represents the query vector; Key vector; Represents a value vector; Indicates transpose; Indicates the feature dimension; S46: Fuse the global dependent feature vectors, that is, perform element-wise summation and obtain global trend features, predict the shield attitude at future moments based on global trend features, and obtain the first prediction result; And the formula for obtaining the first prediction result is: In the formula: Indicates the prediction step size; This represents the set of weight parameters for the first attitude prediction module; Indicates global trend characteristics; This indicates the predicted attitude of the tunnel boring machine at a future moment.
6. The shield tunneling attitude prediction method based on a pseudo-residual hybrid model according to claim 5, characterized in that, The formula for obtaining the second prediction result in S4 is: In the formula: express The first in the model A regression tree; This indicates the corresponding tree structure and leaf node weights; express Preprocessed data and geological parameters of shield tunneling at each moment.
7. The shield tunneling attitude prediction method based on a pseudo-residual hybrid model according to claim 6, characterized in that, The method for obtaining the prediction results of the shield attitude correction at future moments in S4 is as follows: Using the first prediction result as the baseline, the global trend pseudo residual is defined based on the second prediction result; And the expression for the global trend pseudo residual is: In the formula: This indicates the first prediction result; This indicates the second prediction result; This indicates a pseudo-residual representing the global trend. The piecewise mapping strategy is constructed based on the global trend pseudo residual, and its expression is: In the formula: This indicates the predicted attitude correction results for the tunnel boring machine at future moments; , Indicates the adaptive threshold parameter; , This represents the learnable correction coefficient; The shield attitude correction prediction results for future moments are obtained based on the segmented mapping strategy and output through the output module.
8. The shield tunneling attitude prediction method based on a pseudo-residual hybrid model according to claim 7, characterized in that, The method for obtaining the optimal prediction model in S5 is as follows: S51: Use the preprocessed data and shield tunneling geological parameters as sample data, and divide the training set and validation set according to a preset ratio; S52: Train the pseudo-residual mixture prediction model based on the training set to obtain the trained pseudo-residual mixture prediction model; S53: Based on the model loss function, the trained pseudo-residual mixture prediction model is validated using a validation set; That is, to determine whether the output of the pseudo-residual mixture prediction model after training converges; If the output of the trained pseudo-residual fusion prediction model converges, then the trained pseudo-residual fusion prediction model is confirmed to be the optimal prediction model. Otherwise, based on the Bayesian optimization algorithm, the weight parameters of the trained pseudo-residual mixture prediction model are adaptively adjusted, and step S52 is repeated until the weight parameters of the trained pseudo-residual mixture prediction model with converged output are confirmed to be the optimal weight parameters, and the pseudo-residual mixture prediction model is reconstructed to obtain the optimal prediction model.
9. The shield tunneling attitude prediction method based on a pseudo-residual hybrid model according to claim 8, characterized in that, The expression for the model loss function described in S53 is: In the formula: Represents the model loss function; The loss function of the first attitude prediction module is represented, which includes, but is not limited to, the mean squared error function. This represents the loss function output by the dynamic fusion correction module based on global trend pseudo residuals, which includes, but is not limited to, the mean squared error function. This represents the loss function of the second pose prediction module; Represents the loss function term; Represents the regularization term; Indicates the number of controllable leaf nodes; This represents a function term used to control the magnitude of the output value of the leaf nodes; This represents the output value of the j-th leaf node; This represents the weighting coefficient.