A shield tail sealing cavity pressure prediction method based on support vector machine regression

Abstract

A kind of shield tail sealing cavity pressure prediction method based on support vector machine regression belongs to the technical field of tunnel boring machine. According to the operation cycle formed by one-time tunneling of shield machine and one-time segment assembling, multi-source operation data is collected and screened;Using principal component analysis and spearman correlation analysis, the rolling angle of shield, the advancing speed, the outlet pressure of shield tail grease pump 1, the displacement of guide and advancing oil cylinder and the top pressure of slurry are selected as input characteristics;The missing values, abnormal values and pressure time series noise are processed;A regression type support vector machine model is constructed using polynomial kernel, and the model with the smallest five-fold average relative error is selected through grid search and five-fold cross-validation;Pressure prediction and relative error evaluation are carried out. When new operation cycle data is generated, the new data is added to the historical data set, and the original model is replaced when the updated candidate model meets the preset error requirement. The present application can obtain the predicted value of shield tail sealing cavity pressure by using the multi-source operation parameters of shield machine, and realize the rolling update of prediction model.

Description

Technical Field

[0001] This invention belongs to the field of tunnel boring machine technology and relates to a method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression. Specifically, it relates to a method for predicting the pressure of the shield tail sealing cavity using multi-source operating parameters of the tunnel boring machine and a support vector machine regression model. Background Technology

[0002] In urban underground tunnel construction, the shield tunneling method offers advantages such as construction safety, controllable ground settlement, strong adaptability to geological formations, and a high degree of mechanization. During shield tunneling and segment assembly, the tail seal system prevents groundwater, mud, sand, and synchronous grouting fluid from entering the tail of the shield, and its operational status directly affects the safety of shield construction. The pressure in the tail seal cavity reflects the filling and pressurized state of the sealing grease; therefore, reliable monitoring and prediction of this pressure are of great significance.

[0003] Chinese invention patent CN104314580A discloses a tail seal brush and a tunnel boring machine (TBM) for monitoring the pressure of the tail seal grease cavity. This method directly collects the grease cavity pressure by installing a pressure sensor and a wireless signal transmitter within the tail seal brush. While this approach can obtain localized grease cavity pressure, it requires adding sensors and signal transmission components within the tail seal brush. Its technical focus is on direct pressure measurement, without utilizing existing multi-source operating parameters of the TBM to establish a tail seal cavity pressure prediction model. Sun Zhihong and Wang Ning's article, "Experimental Research and Analysis of Tail Seal Leakage Monitoring Based on Fiber Bragg Grating Sensing," published in *Tunnel Construction (Chinese and English)*, Vol. 40, No. 3, 2020, uses fiber Bragg grating sensing to conduct tail seal leakage monitoring experiments. This type of method can obtain monitoring information from specially arranged sensing elements, but it still mainly relies on newly added sensing devices and current measurements, failing to fully utilize the operating data already collected by the TBM during construction, such as propulsion attitude, grease injection, and mud pressure, to establish a pressure mapping relationship.

[0004] The existing technology therefore has the following shortcomings: First, the acquisition of the pressure of the shield tail sealing cavity usually relies on specially arranged pressure or fiber optic sensing elements. When the sensors are affected by complex construction environment and there are missing, sudden changes or noise, the stability of the monitoring results decreases. Second, the existing solutions are mainly for direct pressure monitoring or leakage status identification, and lack the technical process of predicting the pressure of the shield tail sealing cavity from multi-source operating parameters. Third, the tunneling conditions of the tunnel boring machine change with the ring number and stratum, and the fixed model is difficult to continuously adapt to subsequent working conditions.

[0005] Therefore, it is necessary to establish a method for predicting the pressure of the shield tail sealing cavity that can filter, denoise, and select features from the raw operating data and can be updated with new operating cycle data. Summary of the Invention

[0006] To address the problems of existing technologies, such as reliance on a single sensor for obtaining the tail seal cavity pressure, the presence of anomalies and noise in raw operating data, and the insufficient adaptability of fixed prediction models to changes in operating conditions, this invention provides a tail seal cavity pressure prediction method based on support vector machine regression. This invention constructs a multinomial kernel support vector machine regression model by segmenting the multi-source operating parameters of the tunnel boring machine into operating cycles, cleaning the data, performing principal component analysis, and screening Spearman correlations. The model is then optimized and updated using grid search, five-fold cross-validation, and newly added operating cycle data. This establishes a mapping relationship between multi-source operating parameters and the tail seal cavity pressure, enabling tail seal cavity pressure prediction and providing data support for judging the tail seal status and adjusting the grease injection strategy.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression, the method comprising the following steps: Step S1: Collect raw parameter data during the operation of the tunnel boring machine, segment it according to the operating cycle, and perform availability filtering to obtain an available operating cycle dataset; specifically including: Step S1.1: Collect raw parameter data recorded by the tunnel boring machine's sensors; define the process of the tunnel boring machine completing one tunneling and one segment assembly as an operating cycle, and divide the continuously collected raw parameter data into parameter data for multiple operating cycles according to the segment ring number; the raw parameter data includes the tail seal cavity pressure, propulsion speed, shield roll angle, shield pitch angle, tail grease pump outlet pressure, mud top pressure, mud bottom pressure, and guide propulsion cylinder displacement; Step S1.2: Compare the parameter data in each operating cycle with the theoretical parameter range or normal value when the tunnel boring machine is running normally. Remove the operating cycle data corresponding to obvious sensor acquisition abnormalities and abnormal working states of the tunnel boring machine, and retain the operating cycle data with complete time correspondence to form a usable operating cycle dataset.

[0008] Step S2: Based on the available operating cycle data obtained in Step S1, principal component analysis and correlation analysis are performed on the candidate operating parameters to screen out key operating parameters that are highly correlated with the pressure in the shield tail sealing cavity, forming a key input feature set for constructing the pressure prediction model. Specifically: Step S2.1: Principal component analysis is used to perform dimensionality reduction analysis on the candidate operating parameters in the available operating cycle data to reduce information redundancy between candidate operating parameters and obtain the dimensionality reduction analysis results of the candidate operating parameters.

[0009] Step S2.2: Using the shield tail sealing cavity pressure as the target parameter, and combining the dimensionality reduction analysis results obtained in step S2.1, perform Spearman correlation analysis on the candidate operating parameters and the shield tail sealing cavity pressure to obtain the correlation coefficient matrix between each candidate operating parameter and the shield tail sealing cavity pressure.

[0010] Step S2.3: Based on the correlation coefficient matrix, select the original candidate operating parameters with an absolute value of Spearman correlation coefficient not less than 0.4 to obtain the key input feature set; the key input feature set includes shield roll angle, propulsion speed, shield tail grease pump 1 outlet pressure, guide propulsion cylinder displacement and mud top pressure.

[0011] Step S3: Based on the pressure time series of the shield tail sealing cavity obtained in Step S2, perform missing value completion, outlier replacement, and wavelet denoising to obtain a complete denoised and reconstructed pressure feature time series; specifically: Step S3.1: Locate the missing sampling points in the pressure time series of the shield tail sealing cavity, find the two nearest non-missing data before and after the missing sampling point and calculate their mean, and use the mean to fill in the missing values; Step S3.2: Perform abrupt anomaly detection on the pressure data of the shield tail sealing cavity with adjacent sampling points. That is, calculate the mean of the absolute values ​​of the differences between each pressure data and its two adjacent pressure data. If the mean exceeds the set anomaly detection threshold, the abrupt pressure data is determined to be an obvious anomaly and deleted. Then, the missing value mean completion method in step S3.1 is used to complete and replace the deleted data to obtain the processed pressure feature time series. The anomaly detection threshold is 10 bar.

[0012] Step S3.3: Decompose the pressure feature time series obtained in step S3.2 using the five-level wavelet decomposition method. The pressure feature time series is denoted as the zeroth-level approximate sequence. A 0. Using the low-pass and high-pass decomposition filters corresponding to the wavelet basis, approximate and detail components are extracted respectively. Specifically: The first-level wavelet decomposition approximates the zeroth-level sequence. A 0 is decomposed into a first-order approximation coefficient set. A 1 and the first-level detail coefficient group D1; the second-level wavelet decomposition continues to use the first-level approximation coefficient group A 1 is decomposed into a second-order approximation coefficient set. A 2 and second level detail coefficient group D 2; The approximate coefficient group obtained from the previous level is decomposed sequentially until the fifth level wavelet decomposition decomposes the fourth level approximate coefficient group. A 4 is decomposed into a fifth-order approximation coefficient group. A 5 and Level 5 detail coefficient group D 5. After the fifth-level wavelet decomposition is completed, the fifth-level approximation coefficient group is...A5 Recorded as ca 5. Group the detail coefficients from level one to level five. D 1 , D 2 , D 3 , D 4 , D 5 They are respectively denoted as cd 1. cd 2. cd 3. cd 4. cd 5. Finally, a set of approximate coefficients is obtained. ca 5 and five detail coefficient groups cd 1~ cd 5. Among them, ca 5. Used to characterize the overall trend and lowest frequency component of pressure feature time series. cd 1~ cd 5 are used to characterize the detailed changes in components at different decomposition levels, and cd 1 corresponds to the detail component with the highest relative frequency. cd 5 corresponds to relatively low-frequency detail components.

[0013] Step S3.4: Estimate the noise level based on the first-level detail coefficient cd1 and calculate the general threshold. Determine the noise level σ and the threshold λ according to the following equations (1) and (2): (1) (2) In the formula, median(| cd 1|) represents the median of the absolute values ​​of the first-level detail coefficients, and N represents the length of the pressure feature time series.

[0014] Using a combination of hard and soft thresholds to cd 1 to cd 5. The threshold function is expressed as follows (3): (3) In the formula, cd j,k This represents the k-th original detail coefficient at level j. This represents the detail coefficients after processing, sgn represents the sign function, and 0.5 is the adjustment factor.

[0015] Compare the unmodified CA5 with the processed CA5. cd 1 to cd 5. Perform wavelet reconstruction to obtain the denoised and reconstructed time series of pressure features.

[0016] Step S4: Construct a dataset based on the key input feature set obtained in Step S2 and the denoised and reconstructed stress feature time series obtained in Step S3, and establish a regression-type support vector machine model; specifically including: Step S4.1: The input feature matrix is ​​composed of the shield roll angle, propulsion speed, shield tail grease pump 1 outlet pressure, guide propulsion cylinder displacement and mud top pressure. The label vector is composed of the shield tail sealing cavity pressure after denoising reconstruction at the same sampling time. The pressure prediction dataset is constructed together. Normalization is used to eliminate the dimensional differences between the input features to obtain the normalized dataset. Step S4.2: Divide the normalized dataset into two parts: 80% of all samples will be used as the training set and the remaining 20% ​​as the test set. Step S4.3: Construct a regression support vector machine (SVR) model, using a multinomial kernel function, and setting the degree of the multinomial kernel function to 10. The optimization problem of the regression support vector machine is expressed by the following equation (4): (4) In the formula, This represents the weight vector of the regression hyperplane. Indicates the bias term. Indicates the penalty coefficient. Indicates insensitivity to bandwidth loss; and Let i represent the slack variable corresponding to the i-th sample. This represents the deviation where the actual pressure value of the shield tail sealing cavity is higher than the predicted value of the regression function and exceeds the upper boundary of the insensitive zone. The deviation is the amount by which the predicted value of the regression function is higher than the actual pressure value of the shield tail sealing cavity and exceeds the lower boundary of the insensitive zone; This represents the eigenmap determined by the polynomial kernel function; This represents the actual shield tail sealing cavity pressure of the i-th sample; The input feature vector for the i-th sample includes the shield roll angle, propulsion speed, shield tail grease pump 1 outlet pressure, guide propulsion cylinder displacement, and mud top pressure at the same sampling time.

[0017] Step S5: Based on the regression support vector machine model established in Step S4, perform hyperparameter grid search and five-fold cross-validation to obtain the optimal prediction model; specifically: Step S5.1: Obtain the penalty coefficient for the regression support vector machine within the preset search range. Combine the kernel coefficients G of the multinomial kernel function and train the regression support vector machine model corresponding to each hyperparameter combination using the training set; Step S5.2: Divide the training set into five subsets, select one subset as the validation set, and the remaining four subsets as sub-training sets; calculate the mean relative error Eq for each fold and the mean relative error for five folds. E cv Calculate according to equations (5) and (6): (5) (6) In the formula, The average relative error of the validation set in the q-th fold cross-validation; subscript q Indicates the number of cross-validation folds; This represents the number of samples in the q-th fold validation set; p i and Let represent the actual pressure value and the predicted pressure value of the i-th sample in the validation set, respectively; This represents the average relative error of the five-fold cross-validation, i.e., the average relative error obtained from five validations. The arithmetic mean.

[0018] Select the average relative error of five folds The optimal prediction model is obtained by finding the minimum hyperparameter combination and retraining on the entire training set using this hyperparameter combination.

[0019] Step S6: Use the optimal prediction model obtained in Step S5 to perform stress prediction and performance evaluation on the test set, and obtain the stress prediction results for the test set; specifically: Step S6.1: Input the test set input feature vector into the optimal prediction model to obtain the predicted value of the shield tail sealing cavity pressure; input the input feature matrix of the test set obtained in step S4.2 into the optimal prediction model determined in step S5.2 to obtain the predicted value of the shield tail sealing cavity pressure at each sampling time in the test set; wherein, each row of the input feature matrix includes the shield rolling angle, propulsion speed, shield tail grease pump 1 outlet pressure, guide propulsion cylinder displacement and mud top pressure at the same sampling time.

[0020] Step S6.2: Compare the predicted pressure values ​​of the test set with the corresponding actual pressure values. Compare the predicted pressure values ​​of each shield tail sealing cavity obtained in step S6.1 with the actual shield tail sealing cavity pressure values ​​at the corresponding sampling time in the test set. Calculate the relative error of each test sample and the average relative error of the test set to evaluate the prediction performance of the optimal prediction model. Calculate the relative error of a single sample according to equations (7) and (8). e i and the average relative error of the test set ē : (7) (8) In the formula, m represents the number of samples in the test set.

[0021] An average relative error threshold is pre-set based on the allowable pressure prediction error in shield tunneling projects. E max The pressure prediction error allowed by the actual shield tunneling project is predetermined; when ē Not greater than E max When the optimal prediction model is determined to meet the prediction accuracy requirements, it is determined that the optimal prediction model meets the prediction accuracy requirements; when ē Greater than E max If necessary, return to step S5 to adjust the search range of C and G and retrain, validate and evaluate the model.

[0022] Step S7: Receive newly added running cycle data and perform rolling updates when the original prediction model meets the update conditions; specifically: Step S7.1: After the tunnel boring machine completes the excavation and segment assembly of a new ring, extract the original parameter data that is the same as in step S1.1 within the newly added operating cycle; Step S7.2: Perform the data filtering, feature selection, missing value and outlier handling, wavelet denoising and normalization processing of the newly added operating cycle data in sequence from steps S1 to S4, and append the processed newly added operating cycle data to the historical stress prediction dataset to form an updated dataset; Step S7.3: Repeat steps S5 and S6 based on the updated dataset to obtain updated candidate models; when the average relative error of the updated candidate models is not greater than... E max When the optimal prediction model is updated, the candidate model is replaced with the updated one; otherwise, the optimal prediction model is retained. This enables the prediction model to learn and adapt to the changing conditions of the tunnel boring machine.

[0023] The beneficial effects of this invention are as follows: (1) This invention does not directly classify leakage based on the change in the pressure of the shield tail sealing cavity. Instead, it uses the shield rolling angle, propulsion speed, outlet pressure of shield tail grease pump 1, displacement of guide propulsion cylinder and pressure at the top of mud as model inputs, and the shield tail sealing cavity pressure as a continuous prediction target. It uses a regression support vector machine to establish a nonlinear mapping relationship between multi-source operating parameters and shield tail sealing cavity pressure, thereby obtaining the predicted value of shield tail sealing cavity pressure.

[0024] (2) The present invention takes the operating cycle formed by one tunneling and one segment assembly of the shield machine as the data organization and model update unit; when new operating cycle data is generated, the new operating cycle data is combined with historical operating data, and the parameters of the regression support vector machine model are re-optimized and trained, so that the pressure prediction model can be updated in a rolling manner according to the changes in working conditions and geological conditions during the tunneling process.

[0025] (3) Before establishing the pressure prediction model, the present invention performs availability screening, missing value completion, single-point outlier processing and wavelet denoising on the multi-source operating parameters and shield tail sealing cavity pressure data in sequence, so as to reduce the impact of missing, sudden and high-frequency noise of field sensor data on the training results of the pressure prediction model. Attached Figure Description

[0026] Picture 1 This is a schematic diagram of a regression-type support vector machine model. Picture 2 A schematic diagram of nonlinear regressible samples in a low-dimensional space; Picture 3 The Spearman correlation coefficient matrix of the pressure in the tail seal cavity and candidate operating parameters; Picture 4 This is a flowchart of the method of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be further explained below with reference to the operational data and accompanying drawings of a specific tunnel boring machine project.

[0028] The principle of the regression support vector machine model: Regression support vector machine (SVR) is an application of support vector machines in regression problems. Its core principle is to fit the data by minimizing the prediction error. Picture 1 As shown, in SVR, a boundary is defined, consisting of a center line and two parallel boundary lines. The distance between these boundary lines is called the margin, which is usually preset by the user. The goal of SVR is to find a function that minimizes the prediction error of data points that fall outside the margin, and these data points outside the margin are called support vectors. Equation (4) is the objective function and constraints to be optimized in linear soft-margin support vector machine regression. The Lagrange unconstrained optimization function for the above constrained optimization problem is shown in Equation (9): (9) in, a i ≥0, ≥0, μi ≥0, ≥0, i=1,2,3….,N, are called Lagrange multipliers; The Lagrangian function is constructed based on the objective function and constraints of a regression-type support vector machine. For each The Lagrange multiplier vectors formed correspond to the actual pressure being higher than the constraint on the upper boundary of the insensitive zone; For each The Lagrange multiplier vectors formed correspond to the constraint that the actual pressure is lower than the lower boundary of the insensitive zone; For each The Lagrange multiplier vectors formed correspond to Non-negativity constraints; For each The Lagrange multiplier vectors formed correspond to Non-negativity constraints; For vectors The i-th Lagrange multiplier in; Let be the input feature vector of the i-th sample; The actual shield tail sealing cavity pressure value for the i-th sample.

[0029] The original problem of constrained optimization can be expressed as the Lagrange minima problem shown in equation (10): (10) because If the constraint function is a continuously differentiable function and satisfies the KKT conditions, then the primal problem and the dual problem are equivalent. Therefore, the primal problem can be solved by solving the dual problem shown in equation (11). (11) First, by solving the partial derivatives, we obtain the optimization function for... Minimum value: (12) Substitute the four equations in equation (12) above into Can be eliminated This yields the final dual form shown in equation (13).

[0030] (13) in, Let Lagrange multipliers be the insensitive constraints corresponding to the i-th training sample; The kernel function value between the i-th training sample and the j-th training sample; This represents the sequence number of the training sample. The total number of samples used in model training; Taking the negative sign of the objective function and finding its minimum value, we obtain an objective function similar to that of the SVM classification model, as shown in equation (14): (14) By solving the problem shown in formula (14) Then, the regression model coefficients are obtained. , and thus the final linear fitting function is obtained.

[0031] for Picture 2 As shown, for samples exhibiting a nonlinear relationship in low-dimensional space, the linear regression hyperplane cannot adequately describe the mapping relationship between input features and the pressure in the shield tail sealing cavity. Support vector machines utilize kernel functions to calculate the inner product of the high-dimensional feature space in the original space, thereby avoiding the direct construction of a high-dimensional mapping. This implementation uses a polynomial kernel function; commonly used kernel functions are shown in Table 1.

[0032] Table 1: Commonly Used Kernel Functions for Support Vector Machines

[0033] The following is a detailed explanation using a specific example of a tunnel boring machine during its actual advancement process: The shield tail sealing cavity pressure prediction method based on support vector machine regression of the present invention, such as Picture 4 As shown, it includes the following steps: Step S1: Collect raw parameter data during the operation of the tunnel boring machine, segment it according to the operating cycle and perform availability filtering to obtain an available operating cycle dataset.

[0034] Using the operational data of a tunnel boring machine (TBM) in a section of a tunnel project that has completed excavation and segment assembly as the research object, this study extracts data on the tail seal cavity pressure, propulsion speed, shield roll angle, shield pitch angle, tail grease pump outlet pressure, mud top pressure, mud bottom pressure, and guide propulsion cylinder displacement. One excavation and one segment assembly are defined as one operational cycle, and continuous data are segmented according to segment ring number. Based on the theoretical parameter range or normal values ​​of the TBM during normal operation, operational cycles corresponding to obvious sensor anomalies and abnormal TBM operating states are discarded, retaining only usable operational cycle data.

[0035] Step S2: Perform parameter redundancy analysis and correlation screening on the available runtime data to obtain the key input feature set.

[0036] The raw data contained over two hundred operating parameters. Principal component analysis was first performed to identify highly redundant operating parameters using eigenvalues ​​and load coefficients. Then, the Spearman correlation coefficient between the retained original candidate operating parameters and the pressure in the shield tail sealing cavity was calculated. Picture 3This is a correlation coefficient matrix between some candidate operating parameters and the pressure of the shield tail sealing cavity. Each cell in the matrix represents the Spearman correlation coefficient between the corresponding two parameters. Picture 3 The correlation coefficients between the rolling angle of the shield body, the propulsion speed, the outlet pressure of the tail grease pump 1, the displacement of the guide propulsion cylinder, and the pressure at the top of the mud and the pressure in the tail sealing cavity are 0.6449, 0.4445, 0.5643, 0.4202, and 0.603, respectively, all of which are not less than 0.4. Therefore, the five original physical parameters are determined as key input features, and the results are shown in Table 2.

[0037] Table 2: Correlation coefficients between key input features and shield tail sealing cavity pressure

[0038] Step S3: Perform missing value completion, outlier replacement, and wavelet denoising on the key input feature time series and the shield tail sealing cavity pressure time series.

[0039] For missing sampling points, the mean of the two most recent valid data points before and after the missing point is used to fill in the gaps. For abrupt changes in the pressure sequence of the shield tail sealing cavity that occur at only one sampling point and differ significantly from the preceding and following data, an anomaly discrimination factor is calculated. A i ;when A i If the pressure value is greater than 10 bar, delete the pressure value and replace it using the missing value completion method.

[0040] A five-level wavelet decomposition was performed on the pressure feature time series after outlier processing. The first-level decomposition consisted of... ca 0 got ca 1 and cd 1. Continue decomposition at the second level. ca 1. Obtain ca 2 and cd 2. Proceed sequentially to the fifth level, from ca 4. ca 5 and cd 5. Therefore, after the five-level decomposition, a final approximate coefficient is obtained. ca 5 and detail coefficients for five different frequency bands cd 1 to cd 5. use cd 1. Estimate the noise level σ And calculate the threshold λ Meanwhile, the five sets of detail coefficients are calculated, and the adjustment coefficient is set to 0.5. Finally, ca 5 and the processed cd 1 to cd 5. Perform wavelet reconstruction to obtain the denoised and reconstructed time series of pressure features.

[0041] Step S4: Construct a stress prediction dataset and build a regression-type support vector machine model.

[0042] The shield roll angle, propulsion speed, shield tail grease pump 1 outlet pressure, guide propulsion cylinder displacement, and mud top pressure were combined into a five-dimensional input feature vector at the same sampling time. The corresponding denoised and reconstructed shield tail sealing cavity pressure was used as the label value. Table 3 lists some of the normalized five-dimensional input feature data, excluding the shield pitch angle column with a correlation coefficient below 0.4.

[0043] Table 3: Partially Normalized Input Feature Data

[0044] 80% of the normalized dataset was allocated as the training set, and the remaining 20% ​​as the test set. A regression-type support vector machine model was built on the training set, using a polynomial kernel function with a fixed degree of 10.

[0045] Step S5: Perform hyperparameter grid search and five-fold cross-validation to obtain the optimal prediction model.

[0046] The penalty coefficient C and kernel coefficient G are optimized by combination, with candidate levels for both C and G set to [0.0001, 0.0003, 0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1, 3, 10, 30, 100, 1000, 3000, 10000, 30000]. For each combination (C, G), five-fold cross-validation is performed on the training set: the model is trained using four subsets each time, and the mean relative error is calculated using the remaining subset. E q Calculated after five rotations E cv .choose E cv The smallest (C, G) combination is used as the optimal hyperparameter, and the optimal prediction model is obtained by retraining on the entire training set.

[0047] Step S6: Use the optimal prediction model to make predictions on the test set and evaluate the model performance.

[0048] The five-dimensional input feature vector of the test set is input into the optimal prediction model to obtain the corresponding predicted pressure value of the shield tail sealing cavity. Then, the relative error of each test sample and the average relative error of the test set are calculated separately. The average relative error threshold is then determined. E max The pressure prediction error allowed in actual tunnel boring projects is predetermined, and in this embodiment, the average relative error threshold Emax is set to 10%; when the average relative error is not greater than E max When the model meets the prediction accuracy requirements, the average relative error is greater than 1.E max At that time, after adjusting the search ranges of C and G, the grid search, five-fold cross-validation, and model evaluation were performed again.

[0049] Step S7: Update the prediction model using the data from the newly added running cycle.

[0050] After each ring of tunneling and segment assembly is completed by the tunnel boring machine, the original parameter data for that operating cycle is extracted, and the same data processing is performed according to steps S1 to S4. The processed new data is then appended to the historical dataset. Based on the updated dataset, steps S5 and S6 are repeated to form an update candidate model; the update candidate model satisfies... E max If the requirement is met, replace the original optimal prediction model; if the requirement is not met, retain the original model.

[0051] In this way, the prediction model can be continuously adjusted according to the parameter distribution and geological conditions during the tunnel boring machine's excavation process.

[0052] The above embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for predicting the pressure of a shield tail sealing cavity based on support vector machine regression, characterized in that, The method for predicting the pressure in the tail seal cavity includes the following steps: Step S1: Collect raw parameter data during the operation of the tunnel boring machine, segment it according to the operating cycle and perform availability filtering to obtain the available operating cycle dataset; Step S2: Based on available operating cycle data, perform principal component analysis and correlation analysis on candidate operating parameters, screen out key operating parameters that are highly correlated with the pressure of the shield tail sealing cavity, and form a key input feature set for building a pressure prediction model; Step S3: Based on the pressure time series of the shield tail sealing cavity obtained in step S2, perform missing value completion, outlier replacement and wavelet denoising to obtain the complete denoised and reconstructed pressure feature time series; Step S4: Construct a dataset based on the key input feature set obtained in Step S2 and the denoised and reconstructed stress feature time series obtained in Step S3, and establish a regression support vector machine model; Step S5: Based on the regression support vector machine model established in step S4, perform hyperparameter grid search and five-fold cross-validation to obtain the optimal prediction model; Step S6: Use the optimal prediction model obtained in step S5 to perform stress prediction and performance evaluation on the test set, and obtain the stress prediction results for the test set. Step S7: Receive the newly added operating cycle data and perform rolling updates when the original prediction model meets the update conditions, so as to realize the rolling learning and condition adaptation of the prediction model as the shield tunneling working conditions change.

2. The method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression according to claim 1, characterized in that, Step S1 specifically involves: Step S1.1: Collect raw parameter data recorded by the tunnel boring machine's sensors; define the process of the tunnel boring machine completing one tunneling and one segment assembly as an operating cycle, and divide the continuously collected raw parameter data into parameter data for multiple operating cycles according to the segment ring number; the raw parameter data includes the tail seal cavity pressure, propulsion speed, shield roll angle, shield pitch angle, tail grease pump outlet pressure, mud top pressure, mud bottom pressure, and guide propulsion cylinder displacement; Step S1.2: Compare the parameter data in each operating cycle with the theoretical parameter range or normal value when the tunnel boring machine is running normally, remove the operating cycle data corresponding to abnormal sensor acquisition and abnormal working state of the tunnel boring machine, retain the operating cycle data with complete time correspondence, and form a usable operating cycle dataset.

3. The method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression according to claim 2, characterized in that, Step S2 specifically involves: Step S2.1: Principal component analysis is used to perform dimensionality reduction analysis on the candidate operating parameters in the available operating cycle data to reduce information redundancy between candidate operating parameters and obtain the dimensionality reduction analysis results of the candidate operating parameters; Step S2.2: Using the shield tail sealing cavity pressure as the target parameter, and combining the dimensionality reduction analysis results obtained in step S2.1, perform Spearman correlation analysis on the candidate operating parameters and the shield tail sealing cavity pressure to obtain the correlation coefficient matrix between each candidate operating parameter and the shield tail sealing cavity pressure. Step S2.3: Based on the correlation coefficient matrix, select the original candidate operating parameters with an absolute value of Spearman correlation coefficient not less than 0.4 to obtain the key input feature set; the key input feature set includes shield roll angle, propulsion speed, shield tail grease pump 1 outlet pressure, guide propulsion cylinder displacement and mud top pressure.

4. The method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression according to claim 3, characterized in that, Step S3 specifically involves: Step S3.1: Locate the missing sampling points in the pressure time series of the shield tail sealing cavity, find the two nearest non-missing data before and after the missing sampling point and calculate their mean, and use the mean to fill in the missing values; Step S3.2: Perform abrupt anomaly detection on the pressure data of the shield tail sealing cavity with adjacent sampling points. That is, calculate the mean of the absolute values ​​of the differences between each pressure data and its two adjacent pressure data. If the mean value exceeds the set anomaly detection threshold, the abrupt pressure data is determined to be an obvious anomaly and deleted. The missing value mean completion method in step S3.1 is used to complete and replace the deleted data to obtain the processed pressure feature time series. The anomaly detection threshold is 10 bar. Step S3.3: Decompose the pressure feature time series obtained in step S3.2 using the five-level wavelet decomposition method to obtain an approximate coefficient set. ca 5 and five detail coefficient groups cd 1~ cd 5, of which cd 1~ cd 5 represents the first to fifth level detail coefficient groups; Step S3.4: Estimate the noise level based on the first-level detail coefficient cd1 and calculate the general threshold. Determine the noise level σ and the threshold λ according to the following equations (1) and (2): (1) (2) In the formula, median(| cd 1|) represents the median of the absolute values ​​of the first-level detail coefficients, and N represents the length of the pressure feature time series; Using a combination of hard and soft thresholds to cd 1 to cd 5. The threshold function is expressed as follows (3): (3) In the formula, cd j,k This represents the k-th original detail coefficient at level j. This represents the detail coefficients after processing, sgn represents the sign function, and 0.5 is the adjustment factor; Compare the unmodified CA5 with the processed CA5. cd 1 to cd 5. Perform wavelet reconstruction to obtain the denoised and reconstructed time series of pressure features.

5. The method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression according to claim 4, characterized in that, Specifically, step S3.3 is as follows: The pressure characteristic time series is denoted as the zeroth-order approximate series. A 0. Using the low-pass decomposition filter and high-pass decomposition filter corresponding to the wavelet basis, the approximate components and detail components are extracted respectively; The first-level wavelet decomposition approximates the zeroth-level sequence. A 0 is decomposed into a first-order approximation coefficient set. A 1 and the first-level detail coefficient group D1; the second-level wavelet decomposition continues to use the first-level approximation coefficient group A 1 is decomposed into a second-order approximation coefficient set. A 2 and second level detail coefficient group D 2; The approximate coefficient group obtained from the previous level is decomposed sequentially until the fifth level wavelet decomposition decomposes the fourth level approximate coefficient group. A 4 is decomposed into a fifth-order approximation coefficient group. A 5 and Level 5 detail coefficient group D 5; After the fifth-level wavelet decomposition is completed, the fifth-level approximation coefficient group is... A5 Recorded as ca 5. Group the detail coefficients from level one to level five. D 1 , D 2 , D 3 , D 4 , D 5 They are respectively denoted as cd 1. cd 2. cd 3. cd 4. cd 5. Finally, a set of approximate coefficients is obtained. ca 5 and five detail coefficient groups cd 1~ cd 5; among which, ca 5. Used to characterize the overall trend and lowest frequency component of pressure feature time series. cd 1~ cd 5 are used to characterize the detailed changes in components at different decomposition levels.

6. The method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression according to claim 5, characterized in that, Step S4 specifically involves: Step S4.1: The input feature matrix is ​​composed of the shield roll angle, propulsion speed, shield tail grease pump 1 outlet pressure, guide propulsion cylinder displacement and mud top pressure. The label vector is composed of the shield tail sealing cavity pressure after denoising reconstruction at the same sampling time. The pressure prediction dataset is constructed together. Normalization is used to eliminate the dimensional differences between the input features to obtain the normalized dataset. Step S4.2: Divide the normalized dataset into a training set and a test set; Step S4.3: Construct a regression support vector machine model, or SVR model for short, using a multinomial kernel function. The optimization problem of the regression support vector machine is expressed by the following equation (4): (4) In the formula, This represents the weight vector of the regression hyperplane. Indicates the bias term. Indicates the penalty coefficient. Indicates insensitivity to bandwidth loss; and Let i represent the slack variable corresponding to the i-th sample. This represents the deviation where the actual pressure value of the shield tail sealing cavity is higher than the predicted value of the regression function and exceeds the upper boundary of the insensitive zone. The deviation is the amount by which the predicted value of the regression function is higher than the actual pressure value of the shield tail sealing cavity and exceeds the lower boundary of the insensitive zone; This represents the eigenmap determined by the polynomial kernel function; This represents the actual shield tail sealing cavity pressure of the i-th sample; The input feature vector for the i-th sample includes the shield roll angle, propulsion speed, shield tail grease pump 1 outlet pressure, guide propulsion cylinder displacement, and mud top pressure at the same sampling time.

7. The method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression according to claim 6, characterized in that, Step S5 specifically involves: Step S5.1: Combine the penalty coefficient C and kernel coefficient G within the preset search range, and train the regression support vector machine model corresponding to each hyperparameter combination using the training set; Step S5.2: Divide the training set into five subsets, select one subset as the validation set, and the remaining four subsets as sub-training sets; calculate the mean relative error Eq for each fold and the mean relative error for five folds. E cv Calculate according to equations (5) and (6): (5) (6) In the formula, The average relative error of the validation set in the q-th fold cross-validation; subscript q Indicates the number of cross-validation folds; This represents the number of samples in the q-th fold validation set; p i and Let represent the actual pressure value and the predicted pressure value of the i-th sample in the validation set, respectively; This represents the average relative error of the five-fold cross-validation, i.e., the average relative error obtained from five validations. The arithmetic mean; Select the average relative error of five folds E cv The minimum hyperparameter combination is used to retrain on the entire training set to obtain the optimal prediction model.

8. The method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression according to claim 7, characterized in that, Step S6 specifically involves: Step S6.1: Input the test set input feature vector into the optimal prediction model to obtain the predicted value of the shield tail sealing cavity pressure; input the input feature matrix of the test set obtained in step S4.2 into the optimal prediction model determined in step S5.2 to obtain the predicted value of the shield tail sealing cavity pressure at each sampling time in the test set; wherein, each row of the input feature matrix includes the shield rolling angle, propulsion speed, shield tail grease pump 1 outlet pressure, guide propulsion cylinder displacement and mud top pressure at the same sampling time; Step S6.2: Compare the predicted pressure values ​​of the test set with the corresponding actual pressure values. Compare the predicted pressure values ​​of each shield tail sealing cavity obtained in step S6.1 with the actual shield tail sealing cavity pressure values ​​at the corresponding sampling time in the test set. Calculate the relative error of each test sample and the average relative error of the test set to evaluate the prediction performance of the optimal prediction model.

9. The method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression according to claim 8, characterized in that, In step S6.2, the relative error of a single sample is calculated according to equations (7) and (8). e i and the average relative error of the test set ē : (7) (8) In the formula, m represents the number of samples in the test set; An average relative error threshold is pre-set based on the allowable pressure prediction error in shield tunneling projects. E max It is predetermined by the allowable pressure prediction error of the actual shield tunneling project; when ē Not greater than E max When the optimal prediction model is determined to meet the prediction accuracy requirements, it is determined that the optimal prediction model meets the prediction accuracy requirements; when ē Greater than E max If necessary, return to step S5 to adjust the search range of C and G and retrain, validate and evaluate the model.

10. The method for predicting the pressure of the shield tail sealing cavity based on support vector machine regression according to claim 9, characterized in that, Step S7 specifically involves: Step S7.1: After the tunnel boring machine completes the excavation and segment assembly of a new ring, extract the original parameter data that is the same as in step S1.1 within the newly added operating cycle; Step S7.2: Perform the data filtering, feature selection, missing value and outlier handling, wavelet denoising and normalization processing of the newly added operating cycle data in sequence from steps S1 to S4, and append the processed newly added operating cycle data to the historical stress prediction dataset to form an updated dataset; Step S7.3: Repeat steps S5 and S6 based on the updated dataset to obtain updated candidate models; when the average relative error of the updated candidate models is not greater than... E max When the optimal prediction model is updated, the candidate model is replaced with the updated one; otherwise, the optimal prediction model is retained, thus enabling the prediction model to learn and adapt to the changing conditions of the tunnel boring machine.

Images