Shield tunnel muck improvement parameter prediction method and system

Abstract

The invention discloses a shield tunnel muck improvement parameter prediction method and system, and relates to the technical field of tunnel engineering construction.The method comprises the steps that a multi-modal sensor is arranged at a key part of a shield tunneling machine, and tunneling parameters, stratum parameters, multi-modal sensor signals, muck physical properties and modifier injection parameters are collected; then, extracting change characteristics, calculating mutation sensitive factors and carrying out working condition judgment; constructing a normal prediction sub-model library and a sudden change quick response model, calculating a stratum model adaptation index and performing model suitability judgment; establishing a working condition coupling prediction model by using a graph neural network, extracting nonlinear coupling characteristics, calculating a working condition coupling index and judging prediction stability; actual construction performance data are collected and compared with a model prediction result, a prediction correction index is calculated, deviation analysis is carried out, and a corresponding correction strategy is triggered. According to the method, intelligent prediction and dynamic correction of shield tunnel muck improvement parameters can be achieved, and the safety and stability of the construction process are improved.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering construction technology, specifically to a method and system for predicting parameters of soil improvement in shield tunnels. Background Technology

[0002] Shield tunneling technology, due to its minimal surface disturbance, high construction speed, and high safety, has become one of the main construction methods for urban rail transit, underground utility tunnels, and highway tunnels. During shield tunneling, soil preparation is a crucial step in ensuring construction efficiency, safety, and minimizing ground damage, and is therefore of paramount importance. The soil preparation process involves several complex steps, including the amount of conditioner injected, adjustment of tunneling parameters, and determination of ground characteristics, directly affecting the shield tunneling's propulsion efficiency, cutter wear, energy consumption, and ground stability.

[0003] In current shield tunneling construction, the determination of soil improvement parameters largely relies on empirical rules or single sensor data analysis, which presents the following main problems: Incomplete data collection and traditional methods often rely on single-type sensor data, such as torque, thrust, or water content, to make judgments. This ignores the coupling relationship between the multimodal characteristics of the strata and tunneling parameters, resulting in insufficient adaptability to complex working conditions.

[0004] Insufficient feature extraction: Existing technologies mostly rely on raw values ​​or simple statistical analysis for data such as soil moisture content, torque, and earth pressure, lacking in-depth analysis based on time-series change characteristics and abrupt change detection, making it difficult to accurately judge stratum abrupt changes and construction risks.

[0005] The prediction model is singular and has poor adaptability. Existing methods usually use a single prediction model, which lacks the ability to distinguish between normal and sudden operating conditions and has a rapid response mechanism. It is impossible to dynamically adjust the prediction results, which can easily lead to increased prediction bias.

[0006] The lack of model adaptability assessment means that there is no effective model adaptability assessment method in the current prediction of improved parameters for shield tunneling construction. It is impossible to judge the reliability of the prediction model under specific working conditions in real time, and it is difficult to achieve adaptive correction of the model.

[0007] The coupling relationship is not fully utilized. During the tunnel boring machine (TBM) excavation process, there is a complex nonlinear coupling relationship between the tunneling parameters, the stratum parameters and the physical properties of the excavated soil. However, the existing technology has not fully utilized this coupling feature for prediction and optimization.

[0008] Therefore, existing technologies urgently need a method for predicting the parameters of soil improvement in shield tunnels that can comprehensively collect tunneling parameters, geological parameters, multimodal signals, and amendment injection parameters, and combine them with time-series feature extraction, model adaptability assessment, working condition coupling analysis, and online adaptive correction, in order to improve prediction accuracy, enhance model adaptability, increase construction efficiency, and reduce construction risks. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method and system for predicting parameters of soil improvement in shield tunnels, in order to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting parameters of soil improvement in shield tunnels, comprising the following steps: Step 1: By deploying various sensors at key parts of the tunnel boring machine, real-time data are collected on the cutterhead torque, thrust, earth pressure inside the chamber, moisture content of the excavated soil, dry density of the stratum, spectral and resistivity signals, particle size distribution of the excavated soil, liquid and plastic limit parameters of the excavated soil, and parameters for the injection of amendments. Step 2: Based on the moisture content of the slag, cutterhead torque, and resistivity signals, extract the rate of change, acceleration term, and abrupt change amount using time-series difference, sliding window regression, filtered differentiation, and anomaly detection methods, and calculate the abrupt change sensitivity factor. The results are compared with the formation mutation judgment threshold mth to determine whether the current working condition is a normal working condition or a formation mutation working condition. Step 3: A convolutional neural network-based normal prediction sub-model library trained on normal operating condition data is used for prediction of different strata categories; a rapid response model for sudden changes is constructed by combining convolutional neural networks and recurrent neural networks for emergency prediction of sudden operating conditions. The strata model fit index CMI is calculated by fitting historical data and the consistency of strata identification, and compared with the strata model fit threshold Cth to determine whether the current model is suitable. If it is not suitable, the first strategy is given. Step 4: Construct a working condition correlation map based on tunneling parameters, formation parameters and amendment addition amount, establish a working condition coupling prediction model using graph neural network, extract key nonlinear coupling features, calculate the working condition coupling index GCI, and compare it with the working condition coupling threshold Gth to determine whether the prediction is stable. If it is unstable, a second strategy is given. Step 5: Collect actual construction performance data and compare it with the predicted values ​​of the working condition coupled prediction model. Calculate the prediction correction index AEI and compare it with the prediction correction threshold Ath to determine whether the prediction deviation of the working condition coupled prediction model is within a reasonable range. If it is not within a reasonable range, then the third strategy is applied.

[0011] Preferably, step one includes: S11. Collect cutterhead torque data Tp by installing torque sensors on the main drive system of the tunnel boring machine; collect propulsion thrust data F by installing thrust sensors on the propulsion cylinders of the tunnel boring machine; collect earth pressure data P by installing earth pressure sensors on the inner wall of the earth pressure chamber; establish a tunneling parameter data set. S12. Collect soil moisture content data Wt by setting a near-infrared spectral sensor at the soil sampling port; obtain soil dry density data Gd by calling the formation database through the construction log acquisition unit; and establish a formation parameter data group. S13. Collect formation spectral response signal data Snir using a near-infrared spectral sensor; collect formation resistivity signal data Sres using a resistivity probe; establish a multimodal sensor signal data set; S14. Collect the particle size distribution data Dp of the slag soil by arranging a micro-particle online analysis device at the outlet of the soil pressure chamber; collect the liquid limit and plastic limit parameter data Lp of the slag soil by arranging a micro-particle online analysis device at the outlet of the soil pressure chamber; and establish a data set of physical properties of the slag soil. S15. By installing flow sensors, pressure sensors and concentration detection devices on the injection device, the modifier injection parameters, including flow rate, pressure, concentration and injection time, are collected in real time to form a modifier injection parameter set.

[0012] Preferably, step two includes: S21. Based on the moisture content data Wt of the slag soil, the first derivative of the moisture content change in the continuous time series is calculated using time-series difference and sliding window regression techniques, thereby obtaining the rate of change of moisture content. Based on the cutterhead torque data Tp, a time-series filtering and numerical differentiation method is used to first smooth and denoise the torque sequence, and then calculate the second derivative to obtain the acceleration term of the torque. Based on resistivity signal data Sns, anomaly detection and temporal abrupt change analysis methods are employed to compare the differences in resistivity signals between adjacent time periods, identifying significant electrical abrupt changes. The reflectance shift of the spectral response signal data Snir is then used to correct the validity of these resistivity signal abrupt changes, yielding the final resistivity abrupt change amount. .

[0013] Preferably, step two further includes: S22, The rate of change of moisture content obtained acceleration term of torque and resistivity abrupt change After dimensionless processing, the mutation sensitivity factor was calculated and obtained. ; S23. Determine the threshold mth for formation mutation based on a preset formation mutation determination, and set the mutation sensitivity factor... A comparative analysis was performed with the formation abrupt change threshold mth to obtain the first evaluation results, including: When mutation-sensitive factors When the current working condition is ≤ the formation change judgment threshold mth, it is determined to be a normal working condition; When mutation-sensitive factors When the formation mutation determination threshold mth is reached, the current working condition is determined to be a formation mutation working condition.

[0014] Preferably, step three includes: S31. Construct an initial convolutional neural network model using a convolutional neural network. Use the cutterhead torque data Tp, slag moisture content data Wt, resistivity signal data Sres, spectral response signal data Snir, and the corresponding formation type and amendment injection parameters collected under normal operating conditions as training features to train and test the initial convolutional neural network model. Use the trained initial convolutional neural network model as a normal prediction sub-model library, where each sub-model corresponds to a formation type. During runtime, based on the output of the formation identification module, call the prediction sub-model corresponding to the current formation type and output the prediction result. S32. For the sample data of sudden change conditions, a rapid response initial model for sudden change is constructed using a convolutional neural network combined with a recurrent neural network, and the rate of change of water content under sudden change conditions is used as the starting point. acceleration term of torque and resistivity abrupt change As training features, the model is incrementally trained and tested; the trained model is used as a rapid response prediction model for sudden changes. During runtime, when the working condition is determined to be a sudden change, the rapid response model for sudden changes is automatically switched and invoked to generate emergency prediction results. S33. Collect historical construction data and evaluate the goodness of fit between the normal prediction sub-model library and the rapid response model for sudden changes. Compare the sum of squared residuals with the actual collected data to obtain the goodness of fit index Rmodel for each model. S34. Based on the real-time stratigraphic identification results, extract the probability distribution after feature fusion and clustering discrimination, and obtain the consistency probability Pmatch between the stratigraphic identification results and the category of the called prediction model.

[0015] Preferably, step three further includes: S35. After dimensionless processing, the formation model fit index CMI is calculated by using the goodness-of-fit index Rmodel of each model and the consistency probability Pmatch between the formation identification results and the category of the called prediction model. S36. By setting a predefined formation model adaptation threshold Cth, and comparing the formation model adaptation index CMI with the formation model adaptation threshold Cth, the second evaluation results are obtained, including: When the formation model fit index CMI is greater than or equal to the formation model fit threshold Cth, it indicates that the current model is well-fitted and the prediction results output by the current model are maintained. When the formation model fit index (CMI) is less than the formation model fit threshold (Cth), it indicates that the current model is not compatible, triggering the first early warning instruction and generating the first strategy: A multi-model weighting mechanism is adopted, calling pre-trained backup models under different scenarios, and outputting temporary prediction results through weighted averaging to prevent the results from becoming completely invalid; Emergency amendment injection reference parameters for the corresponding formation category are called from the database as protective compensation inputs for the model output, ensuring that even if the model is not compatible, it can still provide a reasonable basis for injection control; The prediction results are dynamically corrected by comparing the model prediction results with the emergency reference values, correcting the prediction results, and forming a temporarily corrected prediction output; While correcting the prediction, the current working condition is marked as "model incompatible state," and key parameters during the correction process, including the CMI value, the corrected prediction value, and the reference value call status, are written to the log.

[0016] Preferably, step four includes: S41. Based on the cutterhead torque data Tp, thrust data F, and internal earth pressure data P from the tunneling parameter group; the slag moisture content data Wt from the formation parameter data group; and the slag particle size distribution data Dp and liquid / plastic limit parameter data Lp from the slag physical property data group, and combined with the actual amount of amendment added, a working condition correlation map is established. An initial working condition coupling model is constructed based on a graph neural network (GNN), using tunneling parameters, formation parameters, and amendment addition amount as input nodes to train and test the initial working condition coupling model. The trained initial working condition coupling model is used as a working condition coupling prediction model to extract nonlinear coupling features during operation, including the correlation between cutterhead torque data and slag moisture content data. Correlation between thrust data and formation pressure distribution data Correlation between in-cabin earth pressure data and spoil permeability data .

[0017] Preferably, step four further includes: S42. Correlation between cutterhead torque data and slag moisture content data, extracted from the nonlinear coupling characteristics of the working condition coupling prediction model. Correlation between thrust data and formation pressure distribution data Correlation between in-cabin earth pressure data and spoil permeability data After dimensionless processing, the working condition coupling index GCI is calculated and obtained. S43. By setting a preset operating condition coupling threshold Gth, and comparing the operating condition coupling index GCI with the operating condition coupling threshold Gth, the third evaluation results are obtained, including: When the working condition coupling index GCI is greater than or equal to the working condition coupling threshold Gth, the prediction is determined to be stable, the output of the working condition coupling prediction model is maintained, and the predicted value of the slag soil improvement parameter is directly generated. When the working condition coupling index GCI is less than the working condition coupling threshold Gth, the prediction is determined to be unstable, triggering a second early warning instruction and generating a second strategy: performing adaptive weighted correction on the input parameters, including adding weight compensation to the torque, pressure, and moisture content signals, and recalculating the prediction results by combining the optimal correction factor from the historical construction database; at the same time, the GCI value, the prediction values ​​before and after correction, and the parameter weighting are recorded in the log for subsequent model retraining and optimization.

[0018] Preferably, step five includes: S51. Collect actual performance data during the construction process, including construction result indicators such as soil flowability Lf, tool wear rate Mc, and energy consumption Ec; at the same time, collect the corresponding parameter values ​​predicted by the working condition coupling prediction model, including predicted values ​​of soil flowability Lpred, tool wear rate Mpred, and energy consumption Epred. S52. Compare the collected actual construction performance data with the predicted parameter values, and calculate the prediction correction index AEI. S53. By setting a preset prediction correction threshold Ath, and comparing the prediction correction index AEI with the prediction correction threshold Ath, the fourth evaluation result is obtained, including: When the prediction correction index AEI is less than the prediction correction threshold Ath, the prediction deviation of the working condition coupled prediction model is determined to be within a reasonable range, no correction is made, and continuous monitoring is carried out. When the prediction correction index AEI is greater than or equal to the prediction correction threshold Ath, it is determined that the prediction deviation of the working condition coupled prediction model is not within a reasonable range and the prediction deviation is large. This triggers the third early warning instruction and generates the third strategy: to start the online correction and transfer learning mechanism, including incrementally updating the parameters of the working condition coupled prediction model and adjusting the model weights and deviations; and to retrain the model in batches based on the AEI values ​​and model correction records recorded in the construction database.

[0019] Preferably, a shield tunnel spoil improvement parameter prediction system includes: The multimodal data acquisition module is used to deploy various sensors in key parts of the tunnel boring machine to collect data in real time, including cutterhead torque, thrust, earth pressure inside the chamber, moisture content of the slag, dry density of the stratum, spectral and resistivity signals, particle size distribution of the slag, liquid and plastic limit parameters of the slag, and parameters for the injection of amendments. The mutation detection and feature extraction module is used to extract the rate of change, acceleration term, and mutation amount based on the moisture content of the slag, cutterhead torque, and resistivity signals, using methods such as time-series difference, sliding window regression, filtered differentiation, and anomaly detection, and to calculate the mutation sensitivity factor. The results are compared with the formation mutation judgment threshold mth to determine whether the current working condition is a normal working condition or a formation mutation working condition. The model prediction and adaptation evaluation module is used for the normal prediction sub-model library of convolutional neural networks trained on normal working condition data, which is used for prediction of different strata categories; the rapid response model for sudden changes is constructed by combining convolutional neural networks and recurrent neural networks, which is used for emergency prediction of sudden working conditions. The strata model fit index CMI is calculated by fitting historical data and the consistency of strata identification, and compared with the strata model fit threshold Cth to determine whether the current model is suitable. If it is not suitable, the first strategy is given. The working condition coupling analysis and prediction module is used to construct a working condition correlation map based on tunneling parameters, formation parameters and amendment addition amount, establish a working condition coupling prediction model using graph neural network, extract key nonlinear coupling features, calculate the working condition coupling index GCI, and compare it with the working condition coupling threshold Gth to determine whether the prediction is stable. If it is unstable, a second strategy is given. The construction performance evaluation and correction module is used to collect actual construction performance data and compare it with the predicted values ​​of the working condition coupling prediction model, calculate the prediction correction index AEI, and compare and analyze it with the prediction correction threshold Ath to determine whether the prediction deviation of the working condition coupling prediction model is within a reasonable range. If it is not within a reasonable range, a third strategy is given.

[0020] This invention provides a method and system for predicting parameters of soil improvement in shield tunnels. It has the following beneficial effects: (1) The method and system for predicting the parameters of soil improvement in shield tunnels use multimodal sensors to collect tunneling parameters, stratum parameters, soil physical properties and improver injection parameters in real time. Combined with time series difference, sliding window regression, filtering differentiation and anomaly detection, the method accurately extracts the rate of change, acceleration term and mutation amount, so as to achieve high-precision real-time prediction of soil improvement parameters in shield tunnels, which significantly improves the reliability and response speed of the prediction.

[0021] (2) The method and system for predicting parameters of soil improvement in shield tunnels introduces a dual-model architecture that combines a normal prediction sub-model library with a rapid response model for sudden changes, and is based on a sudden change sensitive factor. The system can determine the working condition and call different models for normal and sudden working conditions, which improves the system's adaptability and prediction stability under complex working conditions and reduces the construction risks caused by sudden changes in the strata.

[0022] (3) The method and system for predicting parameters of soil improvement in shield tunnels can effectively extract the nonlinear coupling characteristics between tunneling parameters, stratum parameters and improvement agent injection amount by constructing a working condition correlation map and introducing a graph neural network to establish a working condition coupling prediction model, calculate the working condition coupling index GCI, realize the stability judgment and strategy adjustment of complex working conditions, and thus ensure the efficiency, safety and economy of the construction process.

[0023] (4) The method and system for predicting parameters of soil improvement in shield tunnels introduce the prediction correction index AEI and online correction and transfer learning mechanism to realize dynamic comparison, deviation correction and model update between the performance of the construction process and the prediction results, ensure the long-term adaptability and accuracy of the prediction model, and provide technical support for continuous optimization through strategy generation and log recording. Attached Figure Description

[0024] Figure 1 This is a schematic diagram illustrating the steps of a method for predicting parameters for soil improvement in shield tunnels according to the present invention. Figure 2 This is a flowchart of a shield tunnel spoil improvement parameter prediction system according to the present invention. Detailed Implementation

[0025] 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, and 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.

[0026] Example 1 Please see Figure 1 This invention provides a method for predicting parameters of soil improvement in shield tunnels, comprising the following steps: Step 1: By deploying various sensors at key parts of the tunnel boring machine, real-time data are collected on the cutterhead torque, thrust, earth pressure inside the chamber, moisture content of the excavated soil, dry density of the stratum, spectral and resistivity signals, particle size distribution of the excavated soil, liquid and plastic limit parameters of the excavated soil, and parameters for the injection of amendments. Step 2: Based on the moisture content of the slag, cutterhead torque, and resistivity signals, extract the rate of change, acceleration term, and abrupt change amount using time-series difference, sliding window regression, filtered differentiation, and anomaly detection methods, and calculate the abrupt change sensitivity factor. The results are compared with the formation mutation judgment threshold mth to determine whether the current working condition is a normal working condition or a formation mutation working condition. Step 3: A convolutional neural network-based normal prediction sub-model library trained on normal operating condition data is used for prediction of different strata categories; a rapid response model for sudden changes is constructed by combining convolutional neural networks and recurrent neural networks for emergency prediction of sudden operating conditions. The strata model fit index CMI is calculated by fitting historical data and the consistency of strata identification, and compared with the strata model fit threshold Cth to determine whether the current model is suitable. If it is not suitable, the first strategy is given. Step 4: Construct a working condition correlation map based on tunneling parameters, formation parameters and amendment addition amount, establish a working condition coupling prediction model using graph neural network, extract key nonlinear coupling features, calculate the working condition coupling index GCI, and compare it with the working condition coupling threshold Gth to determine whether the prediction is stable. If it is unstable, a second strategy is given. Step 5: Collect actual construction performance data and compare it with the predicted values ​​of the working condition coupled prediction model. Calculate the prediction correction index AEI and compare it with the prediction correction threshold Ath to determine whether the prediction deviation of the working condition coupled prediction model is within a reasonable range. If it is not within a reasonable range, then the third strategy is applied.

[0027] In this embodiment, a method for predicting parameters of soil improvement in shield tunnels is constructed, comprising five system modules: multimodal data acquisition, mutation detection, model prediction and adaptation evaluation, working condition coupling analysis, and construction performance evaluation. This method enables real-time, high-precision prediction and dynamic adaptive adjustment of soil improvement parameters under different geological conditions, significantly improving the safety, stability, and efficiency of shield tunneling while reducing construction risks and operating costs.

[0028] Example 2 This embodiment is an explanation of Embodiment 1. Specifically, step one includes: S11. Collect cutterhead torque data Tp by installing torque sensors on the main drive system of the tunnel boring machine; collect propulsion thrust data F by installing thrust sensors on the propulsion cylinders of the tunnel boring machine; collect earth pressure data P by installing earth pressure sensors on the inner wall of the earth pressure chamber; establish a tunneling parameter data set. S12. Collect soil moisture content data Wt by setting a near-infrared spectral sensor at the soil sampling port; obtain soil dry density data Gd by calling the formation database through the construction log acquisition unit; and establish a formation parameter data group. S13. Collect formation spectral response signal data Snir using a near-infrared spectral sensor; collect formation resistivity signal data Sres using a resistivity probe; establish a multimodal sensor signal data set; S14. Collect the particle size distribution data Dp of the slag soil by arranging a micro-particle online analysis device at the outlet of the soil pressure chamber; collect the liquid limit and plastic limit parameter data Lp of the slag soil by arranging a micro-particle online analysis device at the outlet of the soil pressure chamber; and establish a data set of physical properties of the slag soil. S15. By installing flow sensors, pressure sensors and concentration detection devices on the injection device, the modifier injection parameters, including flow rate, pressure, concentration and injection time, are collected in real time to form a modifier injection parameter set.

[0029] In this embodiment, by deploying multiple sensors at key parts of the tunnel boring machine, real-time data acquisition of cutterhead torque, thrust, earth pressure inside the chamber, moisture content of the excavated soil, dry density of the stratum, spectral and resistivity signals, particle size distribution of the excavated soil, liquid and plastic limit parameters, and amendment injection parameters is achieved. This enables multimodal, all-time, and high-precision data acquisition of tunneling parameters, stratum characteristics, and physical properties of the excavated soil, providing a comprehensive and reliable data foundation for subsequent mutation detection, model prediction, and optimization.

[0030] Example 3 This embodiment is an explanation of embodiment 2. Specifically, step two includes: S21. Based on the moisture content data Wt of the slag soil, the first derivative of the moisture content change in the continuous time series is calculated using time-series difference and sliding window regression techniques, thereby obtaining the rate of change of moisture content. Based on the cutterhead torque data Tp, a time-series filtering and numerical differentiation method is used to first smooth and denoise the torque sequence, and then calculate the second derivative to obtain the acceleration term of the torque. Based on resistivity signal data Sns, anomaly detection and temporal abrupt change analysis methods are employed to compare the differences in resistivity signals between adjacent time periods, identifying significant electrical abrupt changes. The reflectance shift of the spectral response signal data Snir is then used to correct the validity of these resistivity signal abrupt changes, yielding the final resistivity abrupt change amount. .

[0031] In this embodiment, by using methods such as time-series difference, sliding window regression, filtering differentiation, and anomaly detection, the rate of change of soil moisture content, cutterhead torque acceleration term, and resistivity mutation amount are accurately extracted. This enables high-precision identification of the dynamic changes and mutation characteristics of key parameters in the shield tunneling process, providing reliable feature inputs for determining stratum mutations and subsequent model predictions, and improving the timeliness and accuracy of predictions.

[0032] Example 4 This embodiment is an explanation of embodiment 3. Specifically, step two further includes: S22, The rate of change of moisture content obtained acceleration term of torque and resistivity abrupt change After dimensionless processing, the mutation sensitivity factor was calculated and obtained. The formula is as follows:

[0033] In the formula, w1, w2, and w3 represent weighting coefficients; The second derivative of the cutterhead torque characterizes the impact of abrupt changes in formation. It has a high weight and is a key indicator that directly reflects the abrupt changes in torque. This characterizes the impact of changes in soil moisture content gradient on abrupt changes in formation, and has a moderate weight, reflecting the sensitivity of formation moisture content fluctuations and abrupt changes. This characterizes the impact of resistivity signal differences on abrupt changes in formation, and has a secondary weight, reflecting the contribution of formation conductivity differences to abrupt changes in operating conditions. S23. Determine the threshold mth for formation mutation based on a preset formation mutation determination, and set the mutation sensitivity factor... A comparative analysis was performed with the formation abrupt change threshold mth to obtain the first evaluation results, including: When mutation-sensitive factors When the current working condition is ≤ the formation change judgment threshold mth, it is determined to be a normal working condition; When mutation-sensitive factors When the formation mutation determination threshold mth is reached, the current working condition is determined to be a formation mutation working condition.

[0034] The method for obtaining the ground change threshold mth is as follows: Statistical analysis is performed on the temporal variation characteristics of data such as soil moisture content, cutterhead torque, and resistivity signals during a large amount of shield tunneling construction. The distribution range of change-sensitive factors under normal and abrupt change conditions is extracted. Combined with construction experience and geological survey results, a reasonable threshold range is determined. Specific values ​​are determined by referring to shield tunneling industry standards, geological engineering specifications, and relevant construction cases to effectively distinguish between normal and ground change conditions, ensuring the safety and adaptability of the tunneling process.

[0035] In this embodiment, mutation-sensitive factors are constructed. By combining the preset ground change judgment threshold mth, the system enables rapid and accurate determination of normal working conditions and ground change conditions during shield tunneling, which helps to improve the real-time performance and reliability of working condition identification, and thus provides a solid basis for subsequent model switching and emergency response.

[0036] Example 5 This embodiment is an explanation of embodiment 4. Specifically, step three includes: S31. Construct an initial convolutional neural network model using a convolutional neural network. Use the cutterhead torque data Tp, slag moisture content data Wt, resistivity signal data Sres, spectral response signal data Snir, and the corresponding formation type and amendment injection parameters collected under normal operating conditions as training features to train and test the initial convolutional neural network model. Use the trained initial convolutional neural network model as a normal prediction sub-model library, where each sub-model corresponds to a formation type. During runtime, based on the output of the formation identification module, call the prediction sub-model corresponding to the current formation type and output the prediction result. S32. For the sample data of sudden change conditions, a rapid response initial model for sudden change is constructed using a convolutional neural network combined with a recurrent neural network, and the rate of change of water content under sudden change conditions is used as the starting point. acceleration term of torque and resistivity abrupt change As training features, the model is incrementally trained and tested; the trained model is used as a rapid response prediction model for sudden changes. During runtime, when the working condition is determined to be a sudden change, the rapid response model for sudden changes is automatically switched and invoked to generate emergency prediction results. S33. Collect historical construction data and evaluate the goodness of fit between the normal prediction sub-model library and the rapid response model for sudden changes. Compare the sum of squared residuals with the actual collected data to obtain the goodness of fit index Rmodel for each model. S34. Based on the real-time stratigraphic identification results, extract the probability distribution after feature fusion and clustering discrimination, and obtain the consistency probability Pmatch between the stratigraphic identification results and the category of the called prediction model.

[0037] In this embodiment, by constructing a normal prediction sub-model library and a rapid response prediction model for sudden changes, and combining the stratum identification results to perform dynamic model switching, accurate prediction and rapid response of soil improvement parameters under different working conditions are achieved, which greatly improves the adaptability and prediction accuracy of the shield tunneling process.

[0038] Example 6 This embodiment is an explanation of embodiment 5. Specifically, step three further includes: S35. Using the goodness-of-fit index Rmodel and the consistency probability Pmatch between the formation identification results and the category of the called prediction model, after dimensionless processing, the formation model fit index CMI is calculated as follows:

[0039] In the formula, a1 and a2 represent weighting coefficients; The correlation coefficient Rmodel represents the influence of the model fit on the fitness, accounting for the main weight, and reflects the predictive model's ability to interpret the current stratigraphic data; : Characterizes the impact of the stratigraphic matching probability Pmatch on the adaptability, and has a secondary weight, reflecting the degree of conformity between the stratigraphic identification results and the actual working conditions; S36. By setting a predefined formation model adaptation threshold Cth, and comparing the formation model adaptation index CMI with the formation model adaptation threshold Cth, the second evaluation results are obtained, including: When the formation model fit index CMI is greater than or equal to the formation model fit threshold Cth, it indicates that the current model is well-fitted and the prediction results output by the current model are maintained. When the formation model fit index (CMI) is less than the formation model fit threshold (Cth), it indicates that the current model is not compatible, triggering the first early warning instruction and generating the first strategy: A multi-model weighting mechanism is adopted, calling pre-trained backup models under different scenarios, and outputting temporary prediction results through weighted averaging to prevent the results from becoming completely invalid; Emergency amendment injection reference parameters for the corresponding formation category are called from the database as protective compensation inputs for the model output, ensuring that even if the model is not compatible, it can still provide a reasonable basis for injection control; The prediction results are dynamically corrected by comparing the model prediction results with the emergency reference values, correcting the prediction results, and forming a temporarily corrected prediction output; While correcting the prediction, the current working condition is marked as "model incompatible state," and key parameters during the correction process, including the CMI value, the corrected prediction value, and the reference value call status, are written to the log.

[0040] The method for obtaining the stratigraphic model adaptation threshold Cth is as follows: Convolutional neural networks and recurrent neural networks are used to analyze the prediction and fitting of historical construction data for different stratigraphic categories. The goodness-of-fit indices and model consistency probability distributions of various stratigraphic models are statistically analyzed. Combined with the experience of professional technicians and the accuracy requirements for stratigraphic identification, a reasonable threshold range is determined. Referring to industry standards and construction experience for stratigraphic prediction models, this threshold is formulated to ensure a high degree of adaptation between the selected model and actual working conditions, thereby improving prediction accuracy and reducing construction risks.

[0041] In this embodiment, by introducing the formation model fit index CMI and the fit threshold Cth, dynamic evaluation and adaptive correction of the predictive model fit are realized, ensuring that when the model is not fit, the backup model can be switched in time and protective compensation can be performed, effectively improving the stability and reliability of the prediction.

[0042] Example 7 This embodiment is an explanation of embodiment 6. Specifically, step four includes: S41. Based on the cutterhead torque data Tp, thrust data F, and internal earth pressure data P from the tunneling parameter group; the slag moisture content data Wt from the formation parameter data group; and the slag particle size distribution data Dp and liquid / plastic limit parameter data Lp from the slag physical property data group, and combined with the actual amount of amendment added, a working condition correlation map is established. An initial working condition coupling model is constructed based on a graph neural network (GNN), using tunneling parameters, formation parameters, and amendment addition amount as input nodes to train and test the initial working condition coupling model. The trained initial working condition coupling model is used as a working condition coupling prediction model to extract nonlinear coupling features during operation, including the correlation between cutterhead torque data and slag moisture content data. Correlation between thrust data and formation pressure distribution data Correlation between in-cabin earth pressure data and spoil permeability data .

[0043] In this embodiment, by constructing a working condition correlation map and introducing a graph neural network (GNN) for nonlinear coupling feature extraction, the complex correlation between tunneling parameters, formation characteristics and amendment dosage is effectively revealed, significantly improving the accuracy and adaptability of soil improvement parameter prediction.

[0044] Example 8 This embodiment is an explanation of embodiment 7. Specifically, step four further includes: S42. Correlation between cutterhead torque data and slag moisture content data, extracted from the nonlinear coupling characteristics of the working condition coupling prediction model. Correlation between thrust data and formation pressure distribution data Correlation between in-cabin earth pressure data and spoil permeability data After dimensionless processing, the working condition coupling index GCI is calculated and obtained, as shown in the following formula:

[0045] In the formula, s1, s2 and s3 represent weighting coefficients; Characterizing the correlation between propulsion force and cutterhead torque The impact on the coupling of working conditions has a medium weight, reflecting the coupling strength between tunneling dynamics and load; Characterizing the correlation between earth pressure and soil flowability The impact on the coupling of working conditions has a high weight and is a key indicator, reflecting the matching degree between formation pressure and soil characteristics. Characterizing the correlation between the liquid plastic limit of slag and the amount of amendment injected. The impact on the coupling of working conditions has a secondary weight, reflecting the coupling relationship between soil improvement measures and rheological properties; S43. By setting a preset operating condition coupling threshold Gth, and comparing the operating condition coupling index GCI with the operating condition coupling threshold Gth, the third evaluation results are obtained, including: When the working condition coupling index GCI is greater than or equal to the working condition coupling threshold Gth, the prediction is determined to be stable, the output of the working condition coupling prediction model is maintained, and the predicted value of the slag soil improvement parameter is directly generated. When the working condition coupling index GCI is less than the working condition coupling threshold Gth, the prediction is determined to be unstable, triggering a second early warning instruction and generating a second strategy: performing adaptive weighted correction on the input parameters, including adding weight compensation to the torque, pressure, and moisture content signals, and recalculating the prediction results by combining the optimal correction factor from the historical construction database; at the same time, the GCI value, the prediction values ​​before and after correction, and the parameter weighting are recorded in the log for subsequent model retraining and optimization.

[0046] The method for obtaining the coupling threshold Gth under different working conditions is as follows: Historical data on tunneling parameters, geological parameters, and amendment injection volume during shield tunneling are analyzed using a graph neural network to extract the coupling characteristic distribution range under different working conditions. A reasonable threshold is determined by combining expert experience with construction stability requirements. Referring to shield tunneling working condition analysis standards and engineering safety specifications, a threshold is set to effectively assess the stability of the working condition coupling prediction and ensure the reliability of the prediction results during construction.

[0047] In this embodiment, the working condition coupling index GCI is introduced for dynamic stability assessment, and combined with an adaptive weighted correction mechanism, the input parameters and model output can be adjusted in real time when instability is predicted, which effectively improves the reliability and robustness of the prediction of soil improvement parameters and ensures the safety and construction quality of the shield tunneling process.

[0048] Example 9 This embodiment is an explanation of embodiment 8. Specifically, step five includes: S51. Collect actual performance data during the construction process, including construction result indicators such as soil flowability Lf, tool wear rate Mc, and energy consumption Ec; at the same time, collect the corresponding parameter values ​​predicted by the working condition coupling prediction model, including predicted values ​​of soil flowability Lpred, tool wear rate Mpred, and energy consumption Epred. S52. Compare the collected actual construction performance data with the predicted parameter values, and calculate the prediction correction index AEI, as follows:

[0049] In the formula, d1, d2, and d3 represent weighting coefficients; The Lf deviation, which characterizes the fluidity of the excavated soil, has a major weight and is a key parameter that directly reflects the differences in the excavability of the soil during construction. The bias in tool wear rate Mc represents the impact of prediction correction and has a medium weight, reflecting the difference between tool performance and construction energy consumption. : Characterizes the impact of energy consumption Ec deviation on prediction correction, has a minor weight, and reflects the contribution of the difference between overall energy efficiency and prediction model; S53. By setting a preset prediction correction threshold Ath, and comparing the prediction correction index AEI with the prediction correction threshold Ath, the fourth evaluation result is obtained, including: When the prediction correction index AEI is less than the prediction correction threshold Ath, the prediction deviation of the working condition coupled prediction model is determined to be within a reasonable range, no correction is made, and continuous monitoring is carried out. When the prediction correction index AEI is greater than or equal to the prediction correction threshold Ath, it is determined that the prediction deviation of the working condition coupled prediction model is not within a reasonable range and the prediction deviation is large. This triggers the third early warning instruction and generates the third strategy: to start the online correction and transfer learning mechanism, including incrementally updating the parameters of the working condition coupled prediction model and adjusting the model weights and deviations; and to retrain the model in batches based on the AEI values ​​and model correction records recorded in the construction database.

[0050] The prediction correction threshold Ath is obtained by statistically analyzing the deviations between a large amount of actual construction performance data (such as soil flowability, tool wear rate, energy consumption, etc.) and the predicted values ​​to obtain the distribution range of the prediction correction index. Combined with historical construction error experience and model stability requirements, a reasonable threshold range is determined. Referring to the shield tunneling construction quality control specifications and the adaptive correction standard for prediction models, this threshold is set to effectively identify whether the prediction deviation is within a reasonable range, ensuring the accuracy and safety of construction predictions.

[0051] In this embodiment, the prediction correction index (AEI) is introduced to evaluate the output of the working condition coupled prediction model in real time. Combined with online correction and transfer learning mechanisms, the model parameters and weights can be automatically adjusted when the deviation exceeds a reasonable range, so as to realize the adaptive optimization of the prediction model and significantly improve the accuracy and reliability of the shield tunneling process prediction.

[0052] Example 10 Please refer to the following: A prediction system for soil improvement parameters in shield tunnels. Figure 2 ,include: The multimodal data acquisition module is used to deploy various sensors in key parts of the tunnel boring machine to collect data in real time, including cutterhead torque, thrust, earth pressure inside the chamber, moisture content of the slag, dry density of the stratum, spectral and resistivity signals, particle size distribution of the slag, liquid and plastic limit parameters of the slag, and parameters for the injection of amendments. The mutation detection and feature extraction module is used to extract the rate of change, acceleration term, and mutation amount based on the moisture content of the slag, cutterhead torque, and resistivity signals, using methods such as time-series difference, sliding window regression, filtered differentiation, and anomaly detection, and to calculate the mutation sensitivity factor. The results are compared with the formation mutation judgment threshold mth to determine whether the current working condition is a normal working condition or a formation mutation working condition. The model prediction and adaptation evaluation module is used for the normal prediction sub-model library of convolutional neural networks trained on normal working condition data, which is used for prediction of different strata categories; the rapid response model for sudden changes is constructed by combining convolutional neural networks and recurrent neural networks, which is used for emergency prediction of sudden working conditions. The strata model fit index CMI is calculated by fitting historical data and the consistency of strata identification, and compared with the strata model fit threshold Cth to determine whether the current model is suitable. If it is not suitable, the first strategy is given. The working condition coupling analysis and prediction module is used to construct a working condition correlation map based on tunneling parameters, formation parameters and amendment addition amount, establish a working condition coupling prediction model using graph neural network, extract key nonlinear coupling features, calculate the working condition coupling index GCI, and compare it with the working condition coupling threshold Gth to determine whether the prediction is stable. If it is unstable, a second strategy is given. The construction performance evaluation and correction module is used to collect actual construction performance data and compare it with the predicted values ​​of the working condition coupling prediction model, calculate the prediction correction index AEI, and compare and analyze it with the prediction correction threshold Ath to determine whether the prediction deviation of the working condition coupling prediction model is within a reasonable range. If it is not within a reasonable range, a third strategy is given.

[0053] In this embodiment, through a modular design that integrates multimodal data acquisition, mutation detection, model prediction adaptation, working condition coupling analysis, and construction performance correction, the fusion and intelligent processing of multi-source information during shield tunneling construction are realized. This enables real-time identification of geological mutations, dynamic evaluation of model adaptability, prediction of construction working condition coupling stability, and automatic correction when prediction deviations exceed limits, effectively improving construction safety, prediction accuracy, and decision reliability.

[0054] The threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value, it is acceptable.

[0055] The above formulas are all derived from software simulation using a large amount of data, and are selected to be close to the actual values. The coefficients in the formulas are set by those skilled in the art based on the actual situation. The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or changes made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting a shield tunnel spoil improvement parameter, characterized by, Comprise the following steps: Step one, by laying a variety of sensors in the key parts of the shield machine, real-time acquisition of cutter torque, thrust, soil pressure in the cabin, muck moisture content, formation dry density, spectrum and resistivity signal, muck particle size distribution, muck liquid plastic limit parameters and modifier injection parameters; Step two, based on the water content of slag, cutter torque and resistivity signal, through the time difference, sliding window regression, filtering differential and anomaly detection method to extract the rate of change, acceleration term and mutation, calculate the mutation sensitive factor And with the formation mutation judgment threshold mth comparison analysis, determine the current working condition is normal or formation mutation working condition; Step three, based on the normal condition data trained convolutional neural network normal prediction sub model library, for the prediction of different formation categories; Combined with convolutional neural network and recurrent neural network to build a sudden rapid response model for emergency prediction of sudden working conditions, calculate the formation model adaptation index CMI through the consistency of historical data fitting and formation identification, and compare with the formation model adaptation threshold Cth to judge whether the current model is adaptive, if not, give the first strategy; Step four, based on the construction of working condition correlation map of tunneling parameters, formation parameters and modifier addition amount, the working condition coupling prediction model is established by using graph neural network, the key nonlinear coupling features are extracted, the working condition coupling index GCI is calculated, and the comparison analysis is carried out with the working condition coupling threshold Gth to judge whether the prediction is stable, if not, give the second strategy; Step five, compare the actual performance data with the predicted value of working condition coupling prediction model, calculate the prediction correction index AEI, and compare with the prediction correction threshold Ath to judge whether the prediction deviation of working condition coupling prediction model is within the reasonable range, if not, give the third strategy.

2. The method according to claim 1, characterized in that, The step one comprises: S11, install a torque sensor on the main drive system of the shield machine to collect cutter torque data Tp; install a thrust sensor on the thrust cylinder of the shield machine to collect thrust data F; install a soil pressure sensor on the inner wall of the soil pressure chamber to collect soil pressure data P; establish the tunneling parameter data set; S12, set up a near-infrared spectrum sensor at the muck sampling port to collect muck moisture content data Wt; call the formation database through the construction log collection unit to obtain the formation dry density data Gd; establish the formation parameter data set; S13, collect the formation spectrum response signal data Snir through the near-infrared spectrum sensor; collect the formation resistivity signal data Sres through the resistivity probe; establish the multi-modal sensor signal data set; S14, place a micro-particle online analysis device at the outlet of the soil pressure chamber to collect muck particle size distribution data Dp; place a micro-particle online analysis device at the outlet of the soil pressure chamber to collect muck liquid plastic limit parameter data Lp; establish the muck physical property data set; S15, install flow sensor, pressure sensor and concentration detection device on the injection device to real-time collect modifier injection parameters, including flow, pressure, concentration and injection time, form the modifier injection parameter set.

3. The method according to claim 2, wherein, The step two comprises: S21, based on the slag water content data Wt, using time series difference and sliding window regression technique, the change of water content in the continuous time series is calculated by the first order derivative, so as to obtain the change rate of water content ; based on the cutter torque data Tp, using time series filtering and numerical differentiation method, first the torque sequence is smoothed and denoised, and then the second order derivative is calculated to obtain the acceleration term of torque ; based on the resistivity signal data Sres; using anomaly detection and time series mutation analysis method, the difference of resistivity signal between adjacent time periods is compared, the significant electrical mutation is identified, the reflectivity offset of spectral response signal data Snir is combined to correct the effectiveness of resistivity signal mutation, and the final resistivity mutation variable is obtained .

4. The method according to claim 3, characterized in that, The step two further comprises: S22, the change rate of the acquired water content , the acceleration term of the torque , and the resistivity mutation variable , after dimensionless processing, the mutation sensitive factor is calculated ; S23, determine a preset formation mutation determination threshold mth, and obtain a mutation sensitive factor comparing and analyzing the formation mutation determination threshold mth, and obtaining a first evaluation result. when the mutation sensitive factor when the formation mutation determination threshold mthis less than or equal to the formation mutation determination threshold mth, it is determined that the current working condition is a normal working condition. When the mutation sensitive factor When the formation mutation determination threshold mth is reached, it is determined that the current working condition is a formation mutation working condition.

5. The method according to claim 4, wherein, The step three comprises: S31, an initial convolutional neural network model is constructed by using a convolutional neural network, and the torque data Tp of the cutter head, the moisture content data Wt of the spoil, the resistivity signal data Sres, the spectral response signal data Snir, and the corresponding formation category and modifier injection parameters collected under normal conditions are used as training features to train and test the initial convolutional neural network model; the trained initial convolutional neural network model is used as a normal prediction sub-model library, wherein each sub-model corresponds to a formation category; during operation, according to the output result of the formation identification module, the prediction sub-model corresponding to the current formation category is called, and a prediction result is output; S32, for the mutation working condition sample data, using convolutional neural network combined with recurrent neural network to construct mutation fast response initial model, and the change rate of water content under mutation working condition , the acceleration term of torque and the mutation of resistivity as training features, incremental training and testing of the model; the trained model is used as a mutation fast response prediction model, which automatically switches to call the mutation fast response model when the working condition is determined to be a mutation working condition, and generates an emergency prediction result; S33, historical construction data is collected and acquired, and the goodness of fit of the normal prediction sub-model library and the sudden change quick response model is evaluated, the residual sum of squares is compared with the actual collected data, and the goodness of fit index Rmodel of each model is obtained; S34, based on the real-time formation identification result, the probability distribution after feature fusion and clustering discrimination is extracted, and the consistency probability Pmatch between the formation identification result and the called prediction model category is obtained.

6. The method according to claim 5, wherein, The step three further comprises: S35, by obtaining the goodness of fit index Rmodel of each model and the consistency probability Pmatch between the formation identification result and the called prediction model category, after non-dimensional processing, the formation model adaptation index CMI is calculated and obtained; S36, by presetting the formation model adaptation threshold Cth, and comparing and analyzing the formation model adaptation index CMI and the formation model adaptation threshold Cth, a second evaluation result is obtained, including: When the formation model adaptation index CMI is greater than or equal to the formation model adaptation threshold Cth, it indicates that the current model is adapted, and the prediction result output by the current model is maintained; When the formation model adaptation index CMI is less than the formation model adaptation threshold Cth, it indicates that the current model is not adapted, a first warning instruction is triggered, and a first strategy is generated: a multi-model weighting mechanism is used to call backup models pre-trained under different scenarios, a temporary prediction result is output by weighted average, so that the result will not be completely invalid; the emergency modifier injection reference parameters under the corresponding formation category are called from the database and used as the protective compensation input of the model output, so that even if the model is not adapted, reasonable injection control basis can still be provided; the prediction result is dynamically corrected, the model prediction result is compared with the emergency reference value, the prediction result is corrected, and a temporarily corrected prediction output is formed; while correcting the prediction, the current working condition is marked as "model misfit state", and the key parameters in the correction process including the CMI value, the corrected prediction value and the reference value calling situation are written into the log.

7. The method according to claim 6, wherein, The step four comprises: S41, based on the cutterhead torque data Tp, the advancing force data F and the cabin earth pressure data P of the tunneling parameter group, the muck moisture content data Wt of the stratum parameter group, the muck particle size distribution data Dp and the muck liquid-plastic limit parameter data Lp of the muck physical property data group, and combined with the actual modifier addition amount, a working condition correlation atlas is established; based on the graph neural network GNN, a working condition coupling initial model is constructed, and the tunneling parameters, the stratum parameters and the modifier addition amount are taken as input nodes to train and test the working condition coupling initial model; the trained working condition coupling initial model is taken as a working condition coupling prediction model for extracting nonlinear coupling features in runtime, including the correlation between the cutterhead torque data and the muck moisture content data , the correlation between the advancing force data and the stratum pressure distribution data , and the correlation between the cabin earth pressure data and the muck water permeability data .

8. The method according to claim 7, wherein, The step four further comprises: S42, the correlation between the nonlinear coupling features extracted by the working condition coupling prediction model, the cutterhead torque data and the slag water content data , the correlation between the propulsion force data and the formation pressure distribution data and the correlation between the soil pressure data in the cabin and the slag water permeability data After dimensionless processing, the working condition coupling index GCI is calculated S43, by presetting a working condition coupling threshold Gth, and comparing and analyzing the working condition coupling index GCI and the working condition coupling threshold Gth, a third evaluation result is obtained, including: When the working condition coupling index GCI is greater than or equal to the working condition coupling threshold Gth, it is determined that the prediction is stable, the working condition coupling prediction model output result is maintained, and the spoil improvement parameter prediction value is directly generated; When the working condition coupling index GCI is less than the working condition coupling threshold Gth, it is determined that the prediction is unstable, a second early warning instruction is triggered, and a second strategy is generated: performing adaptive weighted correction on the input parameters, including increasing the weight compensation of the torque, pressure, and water content signals, and combining the optimal correction factor of the historical construction database to recalculate the prediction result; at the same time, the GCI value in the correction process, the prediction value before and after correction, and the parameter weighting are recorded into the log for subsequent model retraining and optimization.

9. The method according to claim 8, wherein, The step five comprises: S51, collecting actual performance data in the construction process, including the slag flowability Lf, the tool wear rate Mc, and the energy consumption Ec construction result index; at the same time, collecting the corresponding parameter values predicted by the working condition coupling prediction model, including the slag flowability prediction value Lpred, the tool wear rate prediction value Mpred, and the energy consumption prediction value Epred; S52, comparing the collected actual construction performance data with the predicted parameter values to calculate the prediction correction index AEI; S53, comparing the prediction correction index AEI with the preset prediction correction threshold Ath, and obtaining a fourth evaluation result, including: When the prediction correction index AEI is less than the prediction correction threshold Ath, it is determined that the prediction deviation of the working condition coupling prediction model is within a reasonable range, no correction is made, and continuous monitoring is performed; When the prediction correction index AEI is greater than or equal to the prediction correction threshold Ath, it is determined that the prediction deviation of the working condition coupling prediction model is not within a reasonable range, the prediction deviation is large, a third early warning instruction is triggered, and a third strategy is generated: starting the online correction and transfer learning mechanism, including incrementally updating the parameters of the working condition coupling prediction model, adjusting the model weight and bias, and retraining the model in batches according to the AEI values recorded in the construction database and the model correction records.

10. A system for predicting parameters of improved soil in shield tunneling, applied to the method for predicting parameters of improved soil in shield tunneling according to any one of claims 1-9, characterized in that, Comprise: A multi-modal data acquisition module for laying multiple sensors at key positions of the shield machine, and real-time acquisition of the cutter torque, thrust force, cabin earth pressure, slag water content, dry density of stratum, spectrum and resistivity signal, slag particle size distribution, slag liquid-plastic limit parameter, and modifier injection parameter; The mutation detection and feature extraction module is used for extracting the change rate, acceleration term and mutation quantity based on the slag water content, cutter torque and resistivity signal through the time sequence difference, sliding window regression, filtering differentiation and anomaly detection method, and calculating the mutation sensitive factor and compared with the stratum mutation determination threshold mth to determine whether the current working condition is a normal working condition or a stratum mutation working condition. A model prediction and adaptation evaluation module for a convolutional neural network normal prediction sub-model library trained based on normal working condition data, used for prediction of different stratum categories; a mutation rapid response model constructed by combining convolutional neural network and recurrent neural network, used for emergency prediction of mutation working conditions, a stratum model adaptation index CMI is calculated through consistency calculation of historical data fitting and stratum identification, and compared with a stratum model adaptation threshold Cth to determine whether the current model is adapted, and if not, a first strategy is given; A working condition coupling analysis and prediction module for constructing a working condition correlation map based on the tunneling parameters, stratum parameters, and modifier addition amount, establishing a working condition coupling prediction model using a graph neural network, extracting key nonlinear coupling features, calculating a working condition coupling index GCI, and comparing it with a working condition coupling threshold Gth to determine whether the prediction is stable, and if not, a second strategy is given; The construction performance evaluation and correction module is used for comparing the actual construction performance data with the predicted value of the working condition coupling prediction model, calculating a prediction correction index AEI, and comparing and analyzing the prediction correction threshold Ath to determine whether the prediction deviation of the working condition coupling prediction model is within a reasonable range, and if not, the third strategy is given.

Images