Shield tunneling full deformation prediction method based on artificial intelligence

Abstract

The application relates to the field of tunnel engineering monitoring and construction technology and discloses a shield construction tunnel total deformation prediction method based on artificial intelligence. The method acquires geological sensing data streams, construction operation data streams and a historical deformation case library; the geological sensing data streams and the construction operation data streams are subjected to multi-modal data fusion to generate a construction environment atlas consistent in time and space; a case sequence matched with the current construction environment atlas is searched from the historical deformation case library, feature saliency evaluation is carried out through the case sequence, and a dominant feature group of total deformation prediction is identified; an algorithm optimization engine is constructed based on the dominant feature group; the algorithm optimization engine is activated to perform parallel processing on the construction environment atlas, and an adaptability measurement set of each algorithm is generated; the adaptability measurement set and real-time construction restriction conditions are integrated, a trade-off decision mechanism is used to select an optimal prediction algorithm; a total deformation prediction model is generated, and the model is integrated into a shield construction monitoring platform.

Description

Technical Field

[0001] This invention relates to the field of tunnel engineering monitoring and construction technology, specifically to an artificial intelligence-based method for predicting the total deformation of shield tunnels. Background Technology

[0002] Tunnel deformation can cause ground subsidence, posing a serious threat to the safety of surrounding buildings and underground pipelines. Ground subsidence can lead to building tilting and cracking, affecting the structural safety and normal use of buildings. Underground pipelines such as water, electricity, and gas lines may rupture or deform under the influence of ground subsidence, causing malfunctions in urban infrastructure and greatly inconveniencing residents. In urban construction, there have been instances where ground subsidence caused by tunnel construction has led to cracks in surrounding old buildings, forcing residents to evacuate urgently, which fully demonstrates the severity of tunnel deformation problems.

[0003] Traditional methods for predicting tunnel deformation, such as stochastic medium theory and empirical formulas, have limitations in practical applications. While stochastic medium theory is widely used in calculating displacement and deformation at various points on the ground under different excavation methods in subway tunnels, this theory is not perfect. It cannot be entirely accurate in calculating displacement and deformation at points within the soil and rock mass, and the parameter values ​​require further research and optimization. Empirical formulas are often based on engineering cases and conditions, lacking broad applicability. When encountering complex geological conditions and construction factors, it is difficult to accurately calculate tunnel deformation. Previous studies have mostly analyzed single construction factors. However, shield tunneling is a complex process, and many factors cause soil deformation, including soil loss, frontal thrust, friction between the shield and surrounding soil, grouting pressure, and the shield machine's attitude. Considering only a single factor cannot fully reflect the true soil deformation caused by shield tunneling, leading to significant deviations between predicted results and actual conditions.

[0004] Limitations of computational dimensionality: Most existing methods can only perform two-dimensional planar calculations, making it difficult to accurately describe the three-dimensional soil deformation caused by shield tunneling. Extensive field data shows that soil deformation caused by shield tunneling is three-dimensional, exhibiting significant spatial effects. During tunnel excavation, the soil not only undergoes displacement and deformation in the horizontal and vertical directions but also changes along the tunnel's axial direction. Two-dimensional calculation methods cannot account for these three-dimensional factors, resulting in predictions that do not accurately reflect the full picture of soil deformation and fail to provide comprehensive and accurate guidance for construction.

[0005] Artificial intelligence (AI) technology possesses a powerful ability to handle complex data and nonlinear relationships, enabling rapid analysis and processing of massive amounts of geological sensing data and construction operation data. By establishing appropriate AI models, hidden patterns and relationships between data can be uncovered, thereby more accurately predicting tunnel deformation. Machine learning algorithms can automatically adjust model parameters by learning from historical data, improving prediction accuracy. Deep learning algorithms can automatically extract data features without manual extraction, significantly improving the efficiency and accuracy of data analysis. Introducing AI technology into the field of shield tunnel deformation prediction is expected to overcome the limitations of existing methods, providing more scientific and accurate deformation predictions for tunnel construction, ensuring construction safety and project quality. Summary of the Invention

[0006] The purpose of this invention is to provide an artificial intelligence-based method for predicting the total deformation of shield tunnels, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this invention provides an artificial intelligence-based method for predicting the total deformation of shield tunnels, the method comprising:

[0008] Acquire geological sensor data streams, construction operation data streams, and historical deformation case libraries;

[0009] Multimodal data fusion is performed on the geological sensing data stream and construction operation data stream to generate a spatiotemporally consistent construction environment map;

[0010] Retrieve case sequences that match the current construction environment map from the historical deformation case library, evaluate the saliency of features through the case sequences, and identify the dominant feature group for full deformation prediction;

[0011] Based on the dominant feature group, an algorithm selection engine is constructed, which includes a baseline artificial intelligence algorithm pool and an extended artificial intelligence algorithm group.

[0012] The algorithm optimization engine is activated to process the spatiotemporally consistent construction environment map in parallel, generating an fitness metric set for each algorithm.

[0013] By integrating the aforementioned adaptive metric set with real-time construction constraints, a trade-off decision-making mechanism is adopted to select the optimal prediction algorithm;

[0014] Using the optimal prediction algorithm and dominant feature set, a full deformation prediction model is generated and integrated into the shield tunneling construction monitoring platform.

[0015] Preferably, the step of retrieving a case sequence from the historical deformation case library that matches the current construction environment map, and then evaluating the saliency of the case sequence to identify the dominant feature group for total deformation prediction includes:

[0016] The current construction environment map is analyzed to extract the environmental map feature vector;

[0017] Calculate the similarity matrix between the feature vector of the environmental map and the feature vector of each case in the historical deformation case library;

[0018] Based on the similarity matrix, historical cases with similarity higher than a preset threshold are filtered to form a set of matching cases;

[0019] Perform feature distribution analysis on the set of matching cases and calculate the variance contribution rate of each feature;

[0020] Based on the variance contribution rate, the feature subset with the highest contribution rate is selected as the dominant feature group for full deformation prediction.

[0021] Preferably, the step of performing multimodal data fusion on the geological sensing data stream and the construction operation data stream to generate a spatiotemporally consistent construction environment map includes:

[0022] The geological sensing data stream is timestamped and missing values ​​are imputed to obtain a regular geological data sequence.

[0023] The construction operation data stream is segmented by operation events and outlier removal is performed to obtain a clean construction data sequence;

[0024] The regular geological data sequence and the clean construction data sequence are mapped to a unified spatiotemporal coordinate system, and a continuous environmental data field is generated through interpolation algorithms;

[0025] Feature extraction and dimensional compression are performed on the environmental data field to form a spatiotemporally consistent construction environment map.

[0026] Preferably, the algorithm selection engine comprises a baseline AI algorithm pool and an extended AI algorithm group, including:

[0027] Load predefined benchmark AI algorithms from machine learning libraries to form a benchmark AI algorithm pool;

[0028] An extended group of artificial intelligence algorithms is generated by searching the hyperparameter space using an automated machine learning framework.

[0029] Configure the parallel execution environment of the baseline artificial intelligence algorithm pool and the extended artificial intelligence algorithm group to ensure that each algorithm processes the input data independently;

[0030] Set a standardized interface for the algorithm's output results.

[0031] Preferably, the activation of the algorithm selection engine performs parallel processing on the spatiotemporally consistent construction environment map to generate an fitness metric set for each algorithm, including:

[0032] The spatiotemporally consistent construction environment map is divided into multiple data blocks and assigned to each algorithm for independent prediction.

[0033] Collect the prediction results of each algorithm for each data block, and calculate the error index between the prediction results and the historical true values;

[0034] Statistically analyze the error distribution of each algorithm across all data blocks, and derive accuracy and stability metrics;

[0035] By combining computational resource consumption data, an adaptive metric set is generated for each algorithm.

[0036] Preferably, the error index between the calculated prediction result and the historical true value includes:

[0037] The prediction results for each data block are normalized.

[0038] Extract the actual deformation of the corresponding data block from the historical true values, and calculate the absolute error and relative error;

[0039] Based on time series analysis technology, the dynamic time regularization distance between the predicted results and the actual deformation is calculated.

[0040] The absolute error, relative error, and dynamic time warping distance are combined to form a comprehensive error index.

[0041] Preferably, the step of employing a trade-off decision-making mechanism to select the optimal prediction algorithm includes:

[0042] The fitness metric set is standardized in multiple dimensions to generate a standardized fitness matrix;

[0043] Define constraint functions for real-time construction constraints and apply these constraint functions to the standardized fitness matrix;

[0044] Based on the standardized fitness matrix, the comprehensive utility value of each algorithm is calculated using multi-attribute utility theory, and the algorithm with the highest comprehensive utility value is selected as the best prediction algorithm.

[0045] Preferably, the step of generating a fully deformable prediction model using the optimal prediction algorithm and the dominant feature set includes:

[0046] The dominant feature set is used to select features from the training data to obtain a simplified training set;

[0047] The simplified training set is input into the best prediction algorithm for model training, and the hyperparameters are adjusted to optimize performance.

[0048] Verify the prediction accuracy of the trained model on the test set;

[0049] Export the trained model parameters and structure, and encapsulate them into a fully deformable prediction model.

[0050] Preferably, integrating the full deformation prediction model into the shield tunneling construction monitoring platform includes:

[0051] The fully deformable prediction model is converted into a lightweight inference module to adapt to the computing resources of the monitoring platform;

[0052] Configure the model input interface to receive real-time construction data streams and set the output interface to send prediction results.

[0053] Preferably, the step of searching the hyperparameter space using an automated machine learning framework includes:

[0054] Define the hyperparameter search range and set the search step size;

[0055] A Bayesian optimization algorithm is used to explore hyperparameter combinations, and the performance of each hyperparameter combination on the validation set is recorded. The search direction is dynamically adjusted based on the validation results.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] During tunnel boring machine (TBM) construction, geological conditions and construction operations are key factors affecting tunnel deformation. This method, by acquiring geological sensor data streams and construction operation data streams, can comprehensively collect various information related to TBM construction. The geological sensor data stream contains rich geological information such as stratum lithology, soil physical and mechanical parameters, and groundwater level, which reflects the complexity and diversity of the geological environment in which TBM construction takes place. The construction operation data stream covers construction parameters such as the TBM's advance speed, cutterhead rotation speed, and grouting pressure, which directly affect the process and effect of TBM construction.

[0058] Multimodal data fusion of these two different types of data streams is performed using advanced fusion algorithms. For example, feature-level fusion methods project geological sensing data and construction operation data into a unified feature space for fusion and analysis. Alternatively, decision-level fusion methods analyze and infer the geological sensing data and construction operation data separately, then fuse the inference results from each modality to arrive at a final conclusion. These fusion methods generate a spatiotemporally consistent construction environment map. This map integrates geological and construction operation information in time and space, comprehensively and accurately reflecting the real-time environment of tunnel boring machine (TBM) construction. In a subway TBM construction project, the construction environment map generated through multimodal data fusion clearly shows the changes in TBM construction parameters under different geological conditions and the impact of these changes on the surrounding soil, providing a reliable basis for subsequent tunnel deformation analysis.

[0059] The historical deformation case database is a valuable accumulation of shield tunneling construction experience, containing tunnel deformation cases under various conditions during past shield tunneling operations. These cases record detailed information such as the construction environment, construction parameters, and the final tunnel deformation results. By retrieving case sequences from the historical deformation case database that match the current construction environment map, and using methods such as similarity calculation, the historical cases most similar to the current construction situation can be found. In the shield tunneling construction of a certain river-crossing tunnel, searching the historical deformation case database revealed multiple cases under similar geological conditions and construction parameters, providing important references for the current construction.

[0060] Feature saliency assessment was performed on the matched case sequences, and feature selection algorithms, such as filtering and wrapping feature selection, were used to identify the feature groups that play a dominant role in the overall tunnel deformation. These dominant feature groups may include geological parameters and key construction operation parameters, which are critical factors affecting tunnel deformation. Accurately identifying these key factors can significantly improve the targeting of subsequent tunnel deformation predictions. In the aforementioned river-crossing tunnel construction, feature saliency assessment revealed that the hardness of the strata and the tunnel boring machine's advance speed are the dominant features affecting tunnel deformation. Therefore, in subsequent construction, the changes in these two factors should be closely monitored, and corresponding measures should be taken to control tunnel deformation.

[0061] The algorithm selection engine is a core component of this method, comprising a baseline AI algorithm pool and an extended AI algorithm group. The algorithms in the baseline AI algorithm pool, such as Support Vector Machines (SVMs) and Decision Trees, each have their own characteristics and applicable scope. SVMs perform well in handling small sample sizes and nonlinear problems; Decision Trees can intuitively display the classification and prediction rules of data. The extended AI algorithm group continuously introduces new and more advanced algorithms, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) from deep learning algorithms, to adapt to evolving technological needs.

[0062] The activation algorithm optimization engine performs parallel processing of spatiotemporally consistent construction environment maps. Utilizing the multi-core computing power of the computer, multiple algorithms are run simultaneously to analyze and process the construction environment maps, quickly generating an fitness metric set for each algorithm. This metric set reflects the performance of each algorithm under the current construction environment, including prediction accuracy and computational efficiency. In the construction of a subway tunnel in a certain city, the analysis results of different algorithms on the construction environment map through parallel processing show that, under complex geological conditions, deep learning algorithms have higher accuracy in predicting tunnel deformation, but the computation time is relatively longer; while the traditional support vector machine algorithm, although slightly less accurate, has a faster computation speed.

[0063] By integrating adaptive metrics with real-time construction constraints and employing a trade-off decision-making mechanism, such as a multi-objective optimization algorithm, the optimal prediction algorithm is selected by comprehensively considering both algorithm performance and actual construction needs. In actual construction, constraints such as time and computational resources may exist, necessitating a trade-off between algorithm accuracy and computational efficiency. In the aforementioned subway construction project, based on real-time construction constraints, an algorithm that achieves a good balance between accuracy and computational efficiency was ultimately selected as the optimal prediction algorithm. This improved the accuracy and reliability of tunnel deformation prediction, enabling it to better adapt to different construction conditions and geological situations.

[0064] A total deformation prediction model is generated using the optimal prediction algorithm and dominant feature set. This model can accurately predict the total deformation of the tunnel based on real-time construction environment information and key feature factors. Integrating the total deformation prediction model into the tunnel boring machine (TBM) construction monitoring platform enables real-time monitoring and prediction of tunnel deformation. During TBM construction, the monitoring platform can acquire construction data in real time and input it into the total deformation prediction model. The model quickly calculates the predicted tunnel deformation and displays the results intuitively on the monitoring interface.

[0065] By monitoring and predicting tunnel deformation in real time, the system can issue timely warnings once abnormal deformation is detected. During the construction of a certain tunnel, when the prediction model detected that tunnel deformation exceeded a preset safety threshold, an alarm was immediately issued, reminding construction personnel to take appropriate measures. This provides a scientific basis for construction decisions. Construction personnel can adjust construction parameters in a timely manner based on the prediction results and warning information, such as reducing the tunnel boring machine's advance speed and increasing the grouting volume, to control tunnel deformation and ensure construction safety. Timely warnings and scientific decision-making can also reduce engineering accidents and economic losses caused by deformation. In situations where tunnel collapse or ground subsidence is possible, taking measures in advance can prevent major accidents, save significant repair costs and time, and ensure the smooth progress of the project and the safety of the surrounding environment. Attached Figure Description

[0066] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based shield tunnel full deformation prediction method described in this invention.

[0067] Figure 2 Flowchart for identifying dominant feature groups;

[0068] Figure 3 A flowchart for constructing an algorithm optimization engine;

[0069] Figure 4 A comparative analysis of errors in tunnel deformation prediction algorithms during shield tunneling.

[0070] Figure 5 This is a verification diagram for the accuracy of the prediction model of full deformation of shield tunnel. Detailed Implementation

[0071] 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.

[0072] Please see Figure 1This invention provides an artificial intelligence-based method for predicting the total deformation of shield tunnels. The method includes: acquiring geological sensor data streams, construction operation data streams, and a historical deformation case library. The geological sensor data streams include real-time monitoring information such as ground pressure and soil displacement; the construction operation data streams cover parameters such as shield machine thrust and tunneling speed; and the historical deformation case library stores deformation records and related environmental data from previous construction projects. Multimodal data fusion is performed on the geological sensor data streams and construction operation data streams to eliminate spatiotemporal inconsistencies between data, generating a unified construction environment map. This map represents the geological and operational conditions within the construction area in a grid format. Then, case sequences matching the current construction environment map are retrieved from the historical deformation case library. By analyzing the feature distribution of these cases, the significance of each feature is evaluated, thereby identifying the dominant feature groups affecting the total deformation prediction, such as specific geological parameters or operational indicators. Based on the dominant feature groups, an algorithm selection engine is constructed. This engine includes a set of benchmark artificial intelligence algorithms and an extended algorithm group generated through automatic machine learning. The engine is designed with a parallel processing structure. The algorithm selection engine is activated to perform parallel computation on the construction environment map. Each algorithm independently outputs prediction results, and an adaptive metric set, such as accuracy and stability, is calculated based on these results. The adaptive metric set is integrated with real-time construction constraints, and a trade-off decision mechanism is used to compare the overall performance of each algorithm to select the optimal prediction algorithm. A full deformation prediction model is trained using the optimal prediction algorithm and the dominant feature set, and the model is deployed to the shield tunneling construction monitoring platform to achieve real-time deformation early warning and decision support.

[0073] Example 1: See Figure 2 In practical implementation, the AI-based method for predicting the total deformation of shield tunnels involves retrieving a sequence of cases from a historical deformation case database that matches the current construction environment map. This sequence is then used to assess feature saliency and identify the dominant feature group for total deformation prediction. This process begins with map parsing of the current construction environment map. Map parsing extracts environmental map feature vectors, which contain multi-dimensional data points such as geological attributes, spatial coordinates, and timestamps. A similarity matrix is ​​calculated between the environmental map feature vectors and the feature vectors of each case in the historical deformation case database. This similarity matrix is ​​constructed using Euclidean distance or a cosine similarity algorithm to ensure that the correlation between each historical case and the current environment is quantified. The Euclidean distance formula is as follows:

[0074]

[0075] The specific formula for the cosine similarity algorithm is as follows:

[0076]

[0077] in, Represents the feature vector of the current environment map. With the feature vector of the j-th historical case The Euclidean distance (the smaller the distance, the higher the similarity). represent and The cosine similarity (ranging from -1 to 1, with higher similarity values ​​closer to 1), where m represents the dimension of the feature vector. Represents the current feature vector The kth eigenvalue, Feature vectors representing historical cases The kth eigenvalue.

[0078] Based on the similarity matrix, historical cases with similarity scores higher than a preset threshold are selected to form a set of matching cases. The preset threshold is typically set between 0.7 and 0.9 to balance retrieval accuracy and coverage. Feature distribution analysis is performed on the matching case set, calculating the variance contribution rate of each feature. The variance contribution rate is derived using principal component analysis or variance decomposition techniques, reflecting the importance of each feature in deformation prediction. The specific formula for the variance contribution rate is as follows:

[0079]

[0080] in, The variance contribution rate of the k-th feature. This represents the eigenvalue corresponding to the k-th feature. It represents the sum of the feature values ​​corresponding to all features.

[0081] Based on the variance contribution rate, the feature subset with the highest contribution rate is selected as the dominant feature group for full deformation prediction. The size of the feature subset can be dynamically adjusted according to engineering needs, for example, selecting the top 15% to 30% of features to optimize model efficiency.

[0082] In practical implementation, multimodal data fusion is performed on the geological sensing data stream and the construction operation data stream to generate a spatiotemporally consistent construction environment map. This process first involves timestamp alignment and missing value imputation of the geological sensing data stream. Timestamp alignment uses a global clock synchronization protocol to ensure data consistency, and missing value imputation uses linear interpolation algorithms or spline interpolation methods to fill data gaps, resulting in a regular geological data sequence. The construction operation data stream undergoes operation event segmentation and outlier removal. Operation event segmentation uses sliding window technology to identify continuous operation segments, and outlier removal uses Z-score detection or isolated forest algorithms to remove outliers, resulting in a clean construction data sequence. The regular geological data sequence and the clean construction data sequence are mapped to a unified spatiotemporal coordinate system. This unified spatiotemporal coordinate system adopts a geographic information system framework or a custom grid system, and a continuous environmental data field is generated using inverse distance weighted interpolation algorithms or kriging interpolation methods. Feature extraction and dimensionality compression are performed on the environmental data field. Feature extraction uses wavelet transform or principal component analysis to identify key patterns, while dimensionality compression reduces data complexity through autoencoder neural networks or linear discriminant analysis, ultimately forming a spatiotemporally consistent construction environment map.

[0083] In some embodiments, the map parsing operation involves decomposing the current construction environment map into structured feature vectors. These feature vectors contain numerical and categorical data. Numerical data includes ground pressure values ​​and soil displacement, while categorical data includes geological type codes and operation status labels. The similarity matrix calculation employs multidimensional scaling analysis or dynamic time warping to accommodate the variability of time-series data. The specific formula for dynamic time warping distance is as follows:

[0084]

[0085] in, The dynamic time warping distance represents the time series feature vectors X (length n) and Y (length p) (the smaller the distance, the higher the similarity). The path weight coefficients represent the path weights and satisfy the dynamic programming path constraints. This represents the feature value of X at the i-th time point. This represents the feature value of Y at the j-th time point.

[0086] The selection process for the matching case set includes sorting and thresholding. Sorting is based on descending similarity scores, and thresholding is implemented using hard or soft thresholding functions. Feature distribution analysis uses statistical methods such as histogram analysis or kernel density estimation to assess the variability of each feature in the matching cases. Variance contribution rate calculation is combined with feature importance scoring, which is derived using random forest algorithms or gradient boosting tree models to enhance the robustness of the evaluation. The selection of dominant feature groups considers feature relevance, excluding redundant features using Pearson correlation coefficients or mutual information measures to ensure the independence and representativeness of feature subsets. The multimodal data fusion process emphasizes data quality control; timestamp alignment uses network time protocols to synchronize device clocks, and missing value imputation employs multiple imputation techniques to handle large-scale missing values. Operation event segmentation is based on a rule engine or machine learning classifier. The rule engine defines event boundaries, such as thrust change points, while the machine learning classifier is trained to recognize operation patterns. Outlier removal incorporates domain knowledge rules, which set reasonable value ranges, such as upper and lower thrust limits, to enhance the accuracy of removal. Mapping to a unified spatiotemporal coordinate system involves coordinate transformation algorithms, which employ affine or projective transformations to ensure data spatial consistency. Interpolation algorithm selection considers data sparsity; radial basis function interpolation is used for high data sparsity, while bilinear interpolation is used for dense data. Feature extraction integrates frequency domain analysis, which uses Fast Fourier Transform to extract periodic patterns. Dimensionality compression uses nonlinear methods such as t-SNE or UMAP to preserve data topology.

[0087] Optionally, graph parsing can be extended to multi-scale parsing, which extracts feature vectors at different resolution levels. The coarse scale captures global patterns, while the fine scale captures local details. Similarity matrix calculation can introduce a weighted scheme, assigning importance weights to different features based on domain expert experience or historical data. The matching case set can be preprocessed using clustering, employing the K-means algorithm to group similar cases, reducing computational overhead. Feature distribution analysis can integrate the time dimension, evaluating dynamic feature changes through sliding window analysis. Variance contribution rate calculation can be combined with regularization techniques, such as Lasso regression, to shrink irrelevant features. Dominant feature groups can be dynamically updated, with the update mechanism adjusting feature subsets based on new case inflows. Multimodal data fusion can support real-time streaming processing, and timestamp alignment can introduce an error tolerance mechanism, allowing small time deviations to be corrected through a time window. Missing value imputation can employ generative adversarial networks (GANs), synthesizing realistic data to fill gaps. Operational event segmentation can fuse multi-sensor signals, such as vibration and image data, through convolutional neural networks. Outlier removal can utilize adaptive thresholding, which dynamically adjusts the threshold size based on data distribution. Mapping to a unified spatiotemporal coordinate system can include elevation correction, which uses a digital elevation model to adjust spatial coordinates. Interpolation algorithms can be optimized for parallel computation, leveraging GPUs to accelerate the interpolation process. Feature extraction can incorporate domain-specific feature engineering, constructing composite features such as stress-strain ratio. Dimensionality reduction can preserve interpretability, which is derived from factor analysis to identify physically meaningful features.

[0088] The process of retrieving case sequences from the historical deformation case library relies on an efficient data indexing structure. This structure uses KD-trees or ball trees to accelerate similarity searches and reduce computation time. Feature saliency assessment integrates multi-source feedback, including input from construction personnel or automated scoring, enhancing assessment reliability. Multimodal data fusion emphasizes scalability, which is achieved through distributed storage systems like Hadoop to manage large data volumes and support horizontal scaling. The generation of a spatiotemporally consistent construction environment map ensures consistency for downstream algorithms, such as prediction models, which directly utilize the map structure for inference. The case library stores deformation-related data throughout the entire shield tunneling construction cycle in a standardized structured format. Each case includes four core modules: environmental features, construction parameters, deformation results, and feedback scores. Feature saliency assessment integrates multi-source feedback, including input from construction personnel or automated scoring, enhancing assessment reliability. For example, regarding CASE-2024-001, construction personnel reported that "face pressure control effectively reduced disturbance to the existing tunnel." The automated system gave a high score based on the deformation rate fluctuation coefficient (0.12), and the combination of these two factors confirmed that "soil pressure" was the key feature. Multimodal data fusion emphasizes scalability, which is achieved through distributed storage systems such as Hadoop to manage large data volumes and support horizontal scaling. The multimodal data in the above case, including the geological survey report (PDF), real-time monitoring curves (CSV), and construction images (JPG), are all stored through the Hadoop Distributed File System, with a total data volume of 12TB. Capacity expansion can be achieved by adding new data nodes. The generation of a spatiotemporally consistent construction environment map ensures consistency with downstream algorithms, such as prediction models, which directly utilize the map structure for inference. For example, when a new project faces a scenario of "burial depth of 17.8m + new loess strata + underpass of existing pipelines", the system quickly matches CASE-2024-001 and CASE-2023-047 through KD tree indexing, extracts the dominant features such as "soil pressure control" and "synchronous grouting parameters", provides training samples and initial parameter values ​​for the prediction model, and ensures that the prediction error is less than 0.3mm.

[0089] In practical implementation, the map parsing operation is further refined. The construction of environmental map feature vectors includes a normalization step, which scales features of different dimensions to a uniform range, such as the [0,1] interval. The similarity matrix calculation employs a multi-kernel learning method, combining linear and radial basis kernels to adapt to different feature types. After the formation of the matching case set, case quality is verified, with consistency checks eliminating contradictory records. Feature distribution analysis performs statistical tests, such as ANOVA, to compare differences between groups. Variance contribution rate calculation integrates cross-validation, which segments the data to assess contribution rate stability. After selecting the dominant feature group, feature combinations are optimized, using greedy or genetic algorithms to search for the optimal subset. The multimodal data fusion process strengthens error handling, with error handling mechanisms detecting data stream interruptions and triggering recovery procedures. Timestamp alignment uses high-precision time sources, such as GPS clocks, to reduce synchronization errors. Missing value imputation evaluates imputation quality, with residual analysis monitoring imputation deviations. Operation event segmentation optimizes the window size, which is adaptively adjusted according to the operation cycle. Outlier removal is logged, supporting post-event auditing and model debugging. Coordinate conflicts are handled by mapping to a unified spatiotemporal coordinate system, with the coordinate conflict resolution algorithm prioritizing high-precision coordinate sources. Interpolation algorithms are validated for accuracy through cross-validation comparing different methods. Feature extraction optimizes computational efficiency by reducing dimensionality explosion through feature hashing. Dimensionality compression assesses information loss through quantification of reconstruction error. The process of retrieving case sequences from a historical deformation case library integrates a caching mechanism, storing frequently accessed cases to accelerate subsequent retrieval. Feature saliency evaluation supports incremental learning, updating feature importance online based on new data. Spatiotemporally consistent construction environment maps are stored using a spatiotemporal database, such as PostGIS, for managing multidimensional data queries.

[0090] Example 2: See Figure 3 In practical implementation, the process of building the algorithm selection engine begins with creating a benchmark AI algorithm pool. This pool is implemented by loading predefined algorithms from open-source machine learning libraries. These predefined algorithms include linear regression, decision trees, random forests, support vector machines, and multilayer perceptrons. Each algorithm uses its default configuration parameters during loading to ensure benchmark comparability. An automated machine learning framework is used to search the hyperparameter space to generate an extended AI algorithm group. The hyperparameter search range is defined to cover core parameters such as learning rate, maximum tree depth, number of hidden layers, and activation function type. A search step size is set to control the exploration granularity; for example, the learning rate is explored from 0.001 to 1.0 with a step size of 0.01. A Bayesian optimization algorithm is used to explore hyperparameter combinations. This algorithm establishes a surrogate model of the objective function based on a Gaussian process. The objective function is the algorithm's performance metric on the validation set, such as mean squared error. The specific formula is:

[0091]

[0092] Where MSE represents the mean squared error (a smaller value indicates better model performance), and q represents the number of samples in the validation set. This represents the algorithm's predicted value for the first sample. This represents the true value of the first sample.

[0093] Record the validation set performance for each hyperparameter combination and dynamically adjust the search direction based on the desired improvement of the acquisition function. The specific formula for the desired improvement acquisition function is as follows:

[0094]

[0095] in, This represents the expected improvement value of the hyperparameter combination x. This represents the mean of the Gaussian process model's predictions for x. This represents the objective function value of the current optimal combination of hyperparameters. This represents the standard deviation of the Gaussian process model's prediction for x. The cumulative distribution function representing the standard normal distribution. It represents the probability density function of the standard normal distribution.

[0096] Configure a parallel execution environment for the baseline AI algorithm pool and the extended AI algorithm group. This environment is built using distributed computing frameworks such as Apache Spark or Ray, allocating an independent computing node or container for each algorithm to ensure that each algorithm processes input data independently, avoiding memory contention or data pollution. Set up a standardized interface for the algorithm output results. This standardized interface defines a unified data structure including an array of predicted values, confidence scores, and processing timestamps, facilitating unified calculation and comparative analysis of subsequent fitness metrics.

[0097] In practical implementation, the automatic machine learning framework's search of the hyperparameter space involves iterative optimization loops. During initialization, a small number of hyperparameter combinations are randomly sampled as initial observation points. The performance of these initial observation points on the validation set is evaluated, and the Gaussian process model is updated. In each iteration, the Bayesian optimization algorithm calculates the acquisition function value based on the current Gaussian process model. The acquisition function balances exploring unknown regions with utilizing known good regions, selecting the hyperparameter combination with the largest acquisition function value as the next evaluation point. The validation set performance of each hyperparameter combination is recorded and stored in a performance database, which records metadata such as hyperparameter configuration, validation set loss function value, and training time. The search direction is dynamically adjusted based on the validation results, and the strategy is dynamically adjusted to monitor the performance improvement slope. When there is no significant improvement after several consecutive iterations, the search range is expanded or the acquisition function is switched. Finally, the top 20% of hyperparameter combinations in terms of validation set performance are selected to form an algorithm set complementary to the benchmark AI algorithm pool.

[0098] In some embodiments, the loading process of the benchmark AI algorithm pool integrates a version control mechanism. This mechanism locks a specific version number of the machine learning library to avoid compatibility issues. Predefined algorithms also include traditional statistical models such as the ARIMA time series model. The automated machine learning framework adopts a modular design, allowing for plug-in replacement of optimization algorithms, such as using tree-structured Parsons optimization instead of Bayesian optimization. The definition of the hyperparameter search range supports a conditional space, which handles dependent parameters; for example, the tree depth parameter is searched only when the algorithm type is a tree model. The search step size is set using an adaptive strategy, which adjusts the step size based on parameter sensitivity analysis. Highly sensitive parameters are searched using smaller step sizes for finer-grained searching. The specific formula for parameter sensitivity analysis is:

[0099]

[0100] in, Represents the k-th hyperparameter The sensitivity coefficient (the larger the absolute value, the higher the sensitivity). represents the model loss function, and P represents the set of hyperparameters.

[0101] The Gaussian process model configuration of the Bayesian optimization algorithm uses the Matrn kernel function. The specific formula for the Matrn kernel function is as follows:

[0102]

[0103] in, Representative hyperparameter combination and The kernel function value, Represents the variance parameter. represents the Euclidean distance, and v represents the smoothness parameter (usually set to 3.1 to balance smoothness and flexibility).

[0104] The Matrn kernel function handles non-smooth objective functions and uses a prior mean function to accelerate convergence. Stratified cross-validation is used to evaluate the validation set performance. Stratified cross-validation maintains data distribution consistency and calculates confidence intervals for performance metrics. The confidence interval formula is:

[0105]

[0106] in, Confidence intervals representing performance metrics The sample mean representing the performance metric. represents the quantile of the t-distribution with n-1 degrees of freedom, s represents the sample standard deviation, and n represents the cross-validation fold number.

[0107] In some embodiments, the configuration of the parallel execution environment emphasizes resource isolation, which is achieved through containerization technologies such as Docker, allocating a fixed number of CPU cores and a memory limit to each algorithm instance. The standardized interface is implemented using a protocol buffer format, which defines the algorithm input / output message structure and generates multi-language binding code. The selection of the distributed computing framework considers fault tolerance mechanisms, which monitor node heartbeats and reschedule tasks in case of failure. Independent processing of input data is ensured through a data partitioning strategy, which creates a copy of the input data for each algorithm and sets read-write locks. The unified computation of the adaptability metrics set relies on an aggregation service, which pulls the outputs of each algorithm and batch-calculates accuracy and stability metrics according to predefined formulas.

[0108] Optionally, the benchmark AI algorithm pool can be extended to integrate pre-trained models, including neural network models fine-tuned on similar geological datasets, to improve benchmark performance through transfer learning. Domain knowledge constraints can be introduced to the hyperparameter search range, setting reasonable limits such as a learning rate not exceeding 0.1 to avoid oscillatory divergence. Bayesian optimization algorithms can be mixed with random search, introducing random exploration in the early iterations to escape local optima. The parallel execution environment can support heterogeneous computing, scheduling algorithms to hardware accelerators such as GPUs or FPGAs. Standardized interfaces can add data validation functionality, verifying output format compliance and automatically correcting common errors. The automated machine learning framework can integrate an early stopping mechanism, monitoring the validation set loss curve and terminating training early when performance saturates. Hyperparameter combinatorial exploration can record exploration paths, visualizing these paths to aid in debugging the optimization process and identifying parameter interactions. Validation set performance evaluation can incorporate multi-objective optimization, balancing prediction accuracy and model complexity simultaneously. Dynamically adjusting the search direction can utilize multi-armed search algorithms, adaptively allocating evaluation resources to promising hyperparameter regions. The generation of extended artificial intelligence algorithm ensembles can include ensemble models, which combine multiple basic algorithms through stacking or voting to improve robustness.

[0109] It is understandable that building an algorithm optimization engine requires balancing exploration and utilization. This balance is achieved by adjusting the acquisition function parameters of the Bayesian optimization algorithm, taking into account both algorithm diversity and optimal performance. The parallel execution environment is designed with horizontal scalability in mind, allowing for the dynamic addition of computing nodes to handle large-scale algorithm sets. The establishment of standardized interfaces promotes system decoupling, ensuring that updates to the algorithm pool do not affect the adaptive metric calculation module. Automated hyperparameter search reduces the cost of manual parameter tuning, allowing engineers to focus on business logic rather than parameter debugging. A whitelist mechanism is implemented for loading the benchmark AI algorithm pool, filtering unstable or computationally intensive algorithm variants to ensure the reliability of benchmark algorithms. Predefined algorithm default parameters are pre-optimized using grid search, which tests parameter combinations on standard datasets and selects median performance configurations. The automatic machine learning framework initializes with prior knowledge, injecting this knowledge into the effective hyperparameter range from historical optimization records to accelerate convergence. Checkpoints are set in the iteration cycle of the Bayesian optimization algorithm, saving the optimization state and supporting interruption recovery and multiple experimental comparisons. Resource scheduling in parallel execution environments uses priority queues, which allocate more time slices to computationally intensive algorithms to ensure fairness.

[0110] In its implementation, the generation process of the extended AI algorithm ensemble includes deduplication detection, which compares the parameter distances between newly generated algorithms and existing algorithms, and merges similar algorithms to avoid redundant computation. Dynamic adjustment of the hyperparameter search range is based on real-time monitoring data, which analyzes parameter distribution skew and expands the search for sparse regions. Validation set performance evaluation integrates statistical significance testing, using t-tests to compare algorithm differences and avoid random improvements. Standardized interfaces for algorithm outputs are version-controlled, ensuring backward compatibility and supporting incremental system upgrades. A unified computation module for the adaptability metric optimizes batch processing efficiency, improving computation speed through vectorized operations and memory mapping techniques. The overall implementation of the algorithm selection engine emphasizes reproducibility, achieved by recording random seeds and complete configuration parameters to reproduce experiments. The collaborative work between the baseline AI algorithm pool and the extended AI algorithm ensemble employs hierarchical scheduling, prioritizing the execution of baseline algorithms before scheduling extended algorithms to leverage caching advantages. The search process of the automated machine learning framework supports hot restarts, continuously optimizing from historical best points to address system interruptions. The monitoring and alarm system for the parallel execution environment detects algorithm crashes or resource leaks and triggers a restart mechanism. The resulting algorithm optimization engine outputs a standardized description file, which includes algorithm configuration, performance metrics, and runtime requirements, for downstream decision-making mechanisms to consume.

[0111] Example 3: In specific implementation, the algorithm selection engine is activated to process the spatiotemporally consistent construction environment map in parallel and generate an fitness metric set for each algorithm. This process begins by dividing the spatiotemporally consistent construction environment map into multiple data blocks. The data block division adopts a spatial grid partitioning method, dividing the three-dimensional construction area into fixed-size cubic units. Each data block contains geological sensing data and construction operation data within a continuous time slice. Data blocks are allocated to each algorithm for independent prediction. The allocation mechanism is based on a load balancing strategy, which monitors the real-time load of each computing node and evenly distributes data blocks to each algorithm instance in the baseline AI algorithm pool and the extended AI algorithm group through round-robin scheduling or the least-connection algorithm. The prediction results of each algorithm for each data block are collected. The prediction result collection system establishes a message queue to receive the output of each algorithm instance. The message queue adopts a publish-subscribe model to ensure data integrity and order. The error index between the predicted results and historical actual values ​​is calculated. The error index calculation module extracts the actual deformation of the corresponding data blocks from the historical deformation case library. The actual deformation includes measured data such as tunnel convergence displacement and settlement values. The core error index formulas are the absolute error formula and the relative error formula. The absolute error formula is as follows:

[0112]

[0113] The specific formula for relative error is as follows:

[0114]

[0115] in, This represents the absolute error of the t-th data point. Represents the relative error of the t-th data point (eliminating the influence of dimensions and facilitating comparison across data blocks). Represents the predicted deformation amount. This represents the actual amount of deformation.

[0116] In practice, the prediction results for each data block are normalized to eliminate the influence of dimensions. The normalization process uses a minimum-maximum scaling method to linearly transform the predicted values ​​to the [0,1] interval, as expressed by the formula:

[0117]

[0118] in: This represents the normalized predicted value. Represents the original predicted value. This represents the minimum predicted value within the current data block. This represents the maximum predicted value within the current data block. The actual deformation of the corresponding data block from historical real values ​​is extracted. This extraction is achieved through spatiotemporal coordinate matching, ensuring strict alignment between each predicted point and the measured point. Absolute and relative errors are calculated. The absolute error is the absolute value of the difference between the predicted value and the actual deformation, while the relative error is the ratio of the absolute error to the absolute value of the actual deformation. The dynamic time warping distance between the predicted result and the actual deformation is calculated using time series analysis. This calculation employs a dynamic programming algorithm to find the optimal curvature path between the two time series, minimizing the cumulative distance. The absolute error, relative error, and dynamic time warping distance are aggregated to form a comprehensive error index. The aggregation method uses a weighted summation formula, with weight coefficients allocated according to the importance of the project.

[0119] In practice, the error distribution of each algorithm across all data blocks is statistically analyzed. The error distribution statistics employ kernel density estimation to generate the probability density function of the error values, deriving accuracy and stability metrics. The accuracy metric is calculated as the arithmetic mean of the comprehensive error index across all data blocks, reflecting the overall prediction accuracy of the algorithm. The stability metric is calculated as the standard deviation of the comprehensive error index, characterizing the degree of fluctuation in the algorithm's prediction performance across different data blocks. An adaptive metric set for each algorithm is generated by combining computational resource consumption data, including peak memory usage, CPU time consumption, and disk I / O volume. The adaptive metric set is ultimately organized into a multi-dimensional vector structure.

[0120] In some embodiments, data block partitioning considers spatiotemporal correlation, which is used to determine the optimal block size through variogram analysis, ensuring relatively homogeneous geological conditions within each data block. An independent prediction execution process implements a timeout control mechanism, setting a maximum processing time limit for each algorithm instance. Timeout tasks are marked as failed and rescheduled. A message queue system provides persistent storage, ensuring that prediction results are not lost in the event of system failure. The error index calculation module supports streaming processing, which calculates the error value within a rolling time window in real time.

[0121] In some embodiments, normalization employs robust scaling instead of min-max scaling, using median and quartile ranges to resist outliers. Dynamic time warping distance calculation introduces slope constraints, limiting the inclination angle of curved paths to avoid excessive time alignment distortion. Weight coefficients in the weighted summation formula are determined using the Analytic Hierarchy Process (AHP), which integrates domain expert assessments of the importance of various errors. Computational resource consumption data acquisition integrates operating system-level monitoring, which collects process resource usage in real time via the / proc file system or performance counters. The multidimensional vector of the adaptive metric set contains standardized performance indicators; standardization uses the Z-score method to ensure comparability of indicators with different dimensions. The load balancing strategy introduces performance weights in heterogeneous computing environments, adjusting task allocation ratios based on the algorithm's historical execution efficiency.

[0122] Optionally, data block segmentation can employ adaptive mesh refinement, which densifies the mesh cells in regions with large deformation gradients to improve prediction resolution. Dynamic time warping distance calculation can be combined with derivative dynamic time warping, which considers the first derivative of the time series to improve shape alignment. The generation of the adaptive metric set can introduce an efficiency index, calculated as the number of effective predictions per unit time, measuring algorithm throughput. The parallel processing architecture needs to handle data consistency, which is ensured through a distributed transaction protocol, guaranteeing the atomicity of the processing state of all data blocks. Error distribution statistics employ nonparametric methods, which avoid prior assumptions about the error distribution shape and enhance statistical robustness. It is understood that resource consumption data needs normalization, which converts resource indicators of different dimensions into dimensionless fractions for easier comparison across algorithms.

[0123] In practice, the prediction result collection system implements data verification rules. These rules detect abnormal output values, such as deformation predictions exceeding a reasonable range, and trigger a recalculation process. The actual deformation extraction operation addresses the time offset issue by aligning the predicted and measured time points using an interpolation algorithm. The optimization of the dynamic time warping distance utilizes constraints, which limit the search space for the bending path and reduce computational complexity.

[0124] In practical implementation, error distribution statistics generate a cumulative distribution function, which is used to calculate error percentiles to supplement stability measurement information. The derivation of accuracy and stability metrics simultaneously considers confidence intervals, which are estimated using the Bootstrap resampling method to estimate the statistical uncertainty of the metrics. Data collection for computational resource consumption is filtered to remove instantaneous peaks and prevent occasional fluctuations from affecting the evaluation results. The final organization of the adaptive metric set is stored in a time-series database, which supports querying algorithm performance evolution trends by time range. Monitoring of the parallel processing process is implemented through a visual dashboard, displaying real-time processing progress and resource usage heatmaps for each algorithm. The data block allocation strategy supports priority scheduling, allocating higher computational priority to algorithm instances processing data in critical areas. The error index calculation module integrates anomaly detection, identifying algorithm performance that significantly deviates from the population level and marking it for review. The generation cycle of the adaptive metric set can be adjusted according to the project stage; during high-risk construction phases, the generation cycle is shortened to achieve near real-time algorithm evaluation.

[0125] See Figure 4 This chart, presented in a grouped bar chart format, visually illustrates the error performance of five artificial intelligence algorithms—linear regression, decision tree, random forest, support vector machine, and neural network—in predicting the total deformation of shield tunnels. It compares two core indicators: mean absolute error and root mean square error. This chart provides a quantitative basis for the algorithm selection engine: by clarifying the error boundaries of different algorithms, and considering real-time construction constraints, a trade-off decision mechanism can be used to select the best prediction algorithm. It supports the technical aspects of activating the algorithm selection engine for parallel processing and generating an adaptive metric set, and also provides data support for selecting the total deformation prediction model, ensuring that the model accurately and efficiently outputs deformation warnings on the construction monitoring platform.

[0126] Example 4: In specific implementation, the process of selecting the best prediction algorithm using a trade-off decision mechanism begins with multi-dimensional standardization of the fitness metric set. The fitness metric set includes various performance indicators output from the algorithm optimization engine, such as accuracy metrics, stability metrics, and computational resource consumption data. Multi-dimensional standardization uses a min-max standardization method to linearly transform each metric value to the [0,1] interval, eliminating comparison barriers caused by different dimensions. A standardized fitness matrix is ​​generated, organizing the data in a two-dimensional table format. Rows represent candidate algorithms, columns represent standardized performance indicators, and matrix element values ​​are the standardized scores of each algorithm on the corresponding indicator. Constraint functions for real-time construction limitations are defined, including maximum allowable computational latency, memory usage limits, and power consumption constraints. The constraint functions are expressed as a system of inequalities. These constraint functions are applied to the standardized fitness matrix to filter out algorithm options that do not meet the basic conditions. Based on the standardized fitness matrix, the comprehensive utility value of each algorithm is calculated using multi-attribute utility theory. Multi-attribute utility theory assigns weight coefficients to each performance index, which are determined using the analytic hierarchy process (AHP) or entropy weighting method. The comprehensive utility value is calculated as a weighted arithmetic mean or geometric mean. The algorithm with the highest comprehensive utility value is selected as the best prediction algorithm, and the result is output as an algorithm identifier and a set of configuration parameters.

[0127] In practical implementation, referring to Table 1, the specific operations of multi-dimensional standardization involve data cleaning steps. These steps detect outliers or missing values ​​in the fitness metrics set and use imputation or removal methods to ensure data quality. The min-max standardization formula processes each metric independently, using positive standardization for benefit-type metrics and negative standardization for cost-type metrics. The standardized fitness matrix is ​​constructed and stored in a database table structure for easy subsequent querying and updating. The constraint functions for real-time construction constraints are defined using mathematical expressions; for example, the computational delay constraint is expressed as "algorithm inference time ≤ T_max," and the memory constraint as "algorithm memory usage ≤ M_max." A line-by-line verification mechanism is implemented during the constraint function application process, traversing each row of the standardized fitness matrix and marking algorithm records that violate the constraints. The calculation of multi-attribute utility theory introduces a utility function transformation, which maps standardized scores to decision-maker preference values. The aggregation of comprehensive utility values ​​uses a weighted summation model. The weight coefficient determination process supports dynamic adjustment, updating the weight allocation based on changes in the importance of different construction stages. The final algorithm selection results generate a detailed report, which includes a ranking of the utility values ​​of each algorithm and an explanation of constraint satisfaction.

[0128] Table 1: Standardized Adaptability Matrix

[0129]

[0130] In some embodiments, multi-dimensional standardization processing supports switching between multiple standardization methods, including Z-score standardization and decimal scaling standardization, automatically selecting the appropriate method based on data distribution characteristics. The standardization fitness matrix is ​​stored in a columnar format, which improves query efficiency for a large number of metrics. Constraint functions for real-time construction constraints support logical combination, allowing the use of AND and OR relationships to connect multiple constraints. Multi-attribute utility theory calculations integrate sensitivity analysis, which examines the impact of weight coefficient changes on the ranking results. The weight coefficient determination process introduces a group decision-making mechanism, aggregating weight judgments from multiple domain experts. Multi-dimensional standardization processing can incorporate robust standardization techniques, using median and quartile ranges to resist outlier interference. Soft constraint processing can be implemented in constraint function applications, allowing algorithms with minor constraint violations to proceed to subsequent evaluations but with penalty terms. Multi-attribute utility theory can use the TOPSIS method instead of weighted summation, ranking by calculating proximity to the ideal solution. Weight coefficient determination can be combined with machine learning methods, which learn the optimal weight configuration from historical decision data. A standardized fitness matrix can add a time dimension, recording the changes in indicators during different construction periods to form a three-dimensional evaluation tensor. The constraint function can be dynamically updated, adjusting constraint thresholds in real time based on remaining equipment power or network bandwidth. The calculation of the overall utility value can introduce a nonlinear utility function, which characterizes the nonlinear features of the decision-maker's risk preference.

[0131] In practical implementation, the parameters for multi-dimensional standardization are stored in a configuration file, which records the original minimum and maximum values ​​of each metric for new data standardization. Version management is implemented for the construction of the standardization adaptability matrix, tracking the matrix's modification history to support decision traceability. The constraint function application process records violation details, which help analyze algorithm shortcomings and constraint rationality. The calculation results of the multi-attribute utility theory are normalized, converting the comprehensive utility value to the [0,1] interval for easy comparison. The algorithm selection mechanism implements a review process, with an independent module verifying the correctness of the selection logic. Real-time construction constraints are acquired through a real-time equipment monitoring system, which periodically reports computational resource status data. The constraint function compiler parses textual constraints and generates executable judgment logic code. The multi-attribute utility theory calculation module optimizes vector operation efficiency, using SIMD instructions to process multiple algorithms in parallel. Weight coefficients are stored in a distributed configuration center, which supports synchronous updates of weight parameters across multiple nodes. The final algorithm selection result is pushed to the model deployment system, which loads the corresponding prediction model file based on the algorithm identifier.

[0132] In practice, the entire trade-off decision-making process employs a fault-tolerance mechanism, which uses a backup plan to continue the decision-making process when a single metric data point is abnormal. The standardized fitness matrix is ​​persisted using transactional storage, ensuring the integrity and consistency of the matrix data. A debugging interface is provided during the constraint function application phase, outputting constraint satisfaction details for each algorithm for problem diagnosis. Multi-attribute utility theory calculations support multiple weighting schemes, corresponding to different construction scenario requirements. Algorithm selection results are accompanied by a confidence score, reflecting the reliability of the current decision for reference by subsequent modules.

[0133] Example 5: In specific implementation, the process of generating a full deformation prediction model using the optimal prediction algorithm and dominant feature set is detailed. The dominant feature set is used to select features from the training data, which is a complete dataset from a historical deformation case library. The feature selection operation filters matching fields from all features in the training data based on feature names included in the dominant feature set, such as "soil cohesion," "tunnel boring machine thrust," and "tunneling speed," forming a simplified training set. The simplified training set is then input into the optimal prediction algorithm for model training. The optimal prediction algorithm is assumed to be a gradient boosting decision tree algorithm selected through a trade-off decision mechanism. The model training process uses a mini-batch gradient descent optimizer to adjust hyperparameters to optimize performance. Hyperparameter adjustment involves systematic tuning of parameters such as learning rate, number of trees, and maximum depth. The prediction accuracy of the trained model on the test set is verified. The test set is data from the historical deformation case library, divided into later construction sections according to time sequence, to ensure the model's generalization ability. The model's generalization ability is evaluated by comparing the prediction error distribution on the training and test sets. Export the trained model parameters and structure. Model parameters include the partitioning rules and leaf node values ​​of each tree in the gradient boosting decision tree. The model structure describes the tree connections and feature usage. Package this as a full deformation prediction model, using either the PMLL open standard or a custom binary format. Implement the specific operational process of integrating the full deformation prediction model into the tunnel boring machine (TBM) construction monitoring platform. Convert the full deformation prediction model into a lightweight inference module. The conversion process applies model pruning and quantization techniques. Model pruning removes less important subtrees from the gradient boosting decision tree, and model quantization converts floating-point weights to 8-bit integers to adapt to the monitoring platform's computational resource limitations. Configure the model input interface to receive real-time construction data streams. Define the data format specifications for the model input interface, clearly specifying the field order, data type, and transmission frequency. Set the output interface to send prediction results. Deploy the model service into the monitoring platform's application container during integration. Configure resource quotas and health check endpoints for the application container. Add a prediction result visualization component to the monitoring platform interface. The visualization component displays the tunnel deformation prediction values ​​in the form of curves and color clouds.

[0134] In some embodiments, feature selection operations implement cross-validation to ensure stability. Cross-validation divides the training data into multiple folds, which are used in turn as validation sets to evaluate the performance of feature subsets. A columnar memory layout is used to streamline the storage of the training set, accelerating feature access. An early stopping mechanism is integrated into the model training process. This mechanism monitors the validation set loss curve and terminates training when performance saturates to avoid overfitting. Hyperparameter tuning uses a Bayesian optimization framework, which constructs a Gaussian process model to guide parameter search. Test set partitioning adheres to time-series principles, ensuring that test data timestamps are later than training data to prevent data leakage. Model parameter export includes version metadata, which records training timestamps and data fingerprints for traceability. A lightweight inference module optimizes cache utilization, improving inference speed through feature prefetching and prediction value caching. Data validation is implemented at the model input interface, checking field integrity and numerical rationality. The output interface design supports asynchronous mode, allowing the monitoring platform to obtain prediction results non-blockingly. The application container is configured with an elastic scaling strategy, which dynamically adjusts the number of container instances based on the volume of prediction requests.

[0135] Optionally, feature selection can be combined with recursive feature elimination, which iteratively removes the features with the lowest weights until the target number of features is reached. Model training can employ a distributed training framework, which distributes computational tasks across multiple machines to accelerate the training process. Model validation can incorporate adversarial example testing, injecting slight perturbations to verify model robustness. Model encapsulation can include digital signatures, which verify model integrity and authenticity. The lightweight inference module can support multi-precision inference, dynamically switching computational precision based on platform load to balance speed and accuracy. The output interface can implement a result caching mechanism, storing recent predictions to reduce redundant calculations. The visualization component can integrate an alert function, triggering an audible and visual alarm when a predicted value exceeds a threshold. Container deployment can employ a canary deployment strategy, initially deploying the new model to a small number of nodes to verify stability. The monitoring platform can record model inference logs, including input data, prediction results, and inference time, for performance analysis.

[0136] Optionally, the dominant feature set can be updated periodically, with the update mechanism reassessing feature importance based on newly accumulated construction data. The optimal prediction algorithm can support online learning, with the online learning mechanism using real-time data increments to update model parameters. The full deformation prediction model can be exported to ONNX format, enabling cross-platform deployment compatibility. The lightweight inference module can integrate model interpretation functionality, which outputs feature contribution values ​​to enhance prediction interpretability. The tunnel boring machine (TBM) construction monitoring platform can be configured for A / B testing, comparing the performance differences of different algorithm models in actual construction.

[0137] Understandably, feature selection reduces data dimensionality and improves training efficiency, which in turn allows the model to be deployed to the construction site more quickly. Lightweight model transformation ensures stable operation on resource-constrained monitoring platforms, a fundamental requirement for real-time prediction. Standardized input / output interface design reduces system integration complexity, which in turn facilitates maintenance and expansion by engineers.

[0138] In practical implementation, the feature selection operation example processes a historical dataset containing 200 features. The dominant feature group includes 15 key features such as "cutterhead torque," "soil chamber pressure," and "grouting volume." After feature selection, the dimensionality of the training set is reduced by 92.5%. Model training is performed on a server equipped with GPUs. The gradient boosting decision tree algorithm is set to 1000 trees, and the learning rate is determined to be 0.05 through grid search. Training takes approximately 3 hours. Model validation shows that the mean absolute error on the test set is 0.12 mm, which is close to the training set error of 0.11 mm. The model export file size is compressed to 15MB and contains complete tree structure serialized data. The lightweight inference module is deployed on the edge computing device of the monitoring platform, with 4GB of memory. After the module starts, the memory usage is stable at 800MB. The input interface receives real-time sensor data streams every 5 seconds. After verification, the data streams are converted into 224-dimensional input vectors required by the model. The output interface pushes the predicted values ​​to the message queue of the monitoring platform, and the message queue consumer updates the interface display components. The visualization component displays the tunnel deformation prediction curve for the next 10 minutes at a refresh rate of 30 seconds, with the curve color gradually changing from green to red according to the severity of the prediction.

[0139] See Figure 5This graph, using a double-line comparison, visually presents the dynamic fitting relationship between the number of construction days and actual and predicted deformation. It is a core visualization achievement of the full deformation prediction model generation and verification process. The solid line represents the actual tunnel deformation, and the dashed line represents the model's predicted deformation. The overall trends of the two are highly consistent, especially in the later stages of construction, where the fitting degree is extremely high. This clearly demonstrates the excellent generalization ability of the model trained based on the optimal prediction algorithm and dominant feature set. From a technical perspective, this graph verifies the effectiveness of the entire process: using dominant feature set to simplify the training set, inputting the optimal algorithm to train the model, and verifying the accuracy of the test set. The model not only accurately captures the cumulative growth pattern of tunnel deformation as construction progresses but also maintains consistency between prediction and reality at different construction stages. This provides a solid accuracy guarantee for the subsequent integration of the model into the shield tunneling monitoring platform and the realization of real-time deformation early warning. Simultaneously, this high-precision fitting also indirectly confirms the technical rationality of the previous multimodal data fusion, dominant feature set identification, and algorithm optimization stages. It represents a key achievement of the AI-based shield tunnel full deformation prediction method from data processing to model implementation, directly supporting the invention goal of generating a full deformation prediction model and integrating it into the monitoring platform.

[0140] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A shield tunneling whole deformation prediction method based on artificial intelligence, characterized in that, The method comprises: Obtaining geological sensing data stream, construction operation data stream and historical deformation case library; Multi-modal data fusion is performed on the geological sensing data stream and the construction operation data stream to generate a spatio-temporally consistent construction environment graph; From the historical deformation case library, a case sequence matching the current construction environment graph is retrieved, feature saliency evaluation is performed through the case sequence, and a dominant feature group for full deformation prediction is identified; Based on the dominant feature group, an algorithm optimization engine is constructed, which includes a benchmark artificial intelligence algorithm pool and an extended artificial intelligence algorithm group; The algorithm optimization engine is activated to perform parallel processing on the spatio-temporally consistent construction environment graph, generating an adaptability metric set for each algorithm; The adaptability metric set is integrated with real-time construction restriction conditions, and a trade-off decision mechanism is used to select the best prediction algorithm; Using the best prediction algorithm and the dominant feature group, a full deformation prediction model is generated, and the full deformation prediction model is integrated into a shield construction monitoring platform; The retrieval of the case sequence matching the current construction environment graph from the historical deformation case library, the feature saliency evaluation through the case sequence, and the identification of the dominant feature group for full deformation prediction comprise: The current construction environment graph is analyzed to extract an environment graph feature vector; The similarity matrix between the environment graph feature vector and the feature vector of each case in the historical deformation case library is calculated; According to the similarity matrix, historical cases with a similarity higher than a preset threshold are screened to form a matching case set; Feature distribution analysis is performed on the matching case set, and the variance contribution rate of each feature is calculated; Based on the variance contribution rate, the feature subset with the highest contribution rate is selected as the dominant feature group for full deformation prediction; The activation of the algorithm optimization engine to perform parallel processing on the spatio-temporally consistent construction environment graph to generate an adaptability metric set for each algorithm comprises: The spatio-temporally consistent construction environment graph is divided into multiple data blocks and assigned to each algorithm for independent prediction; The prediction results of each algorithm for each data block are collected, and the error indicators between the prediction results and the historical true values are calculated; The error distribution of each algorithm on all data blocks is counted to derive accuracy metrics and stability metrics; Combined with the calculation resource consumption data, the adaptability metrics of each algorithm are generated. 2.The shield tunneling full deformation prediction method based on artificial intelligence according to claim 1, wherein, The multi-modal data fusion of the geological sensing data stream and the construction operation data stream to generate a spatio-temporally consistent construction environment graph comprises: Time stamp alignment and missing value interpolation are performed on the geological sensing data stream to obtain a regularized geological data sequence; Operation event segmentation and outlier removal are performed on the construction operation data stream to obtain a pure construction data sequence; The regularized geological data sequence and the pure construction data sequence are mapped to a unified spatio-temporal coordinate system, and a continuous environment data field is generated through an interpolation algorithm; Feature extraction and dimension compression are performed on the environment data field to form a spatio-temporally consistent construction environment graph. 3.The shield tunneling full deformation prediction method based on artificial intelligence according to claim 1, wherein, The construction of the algorithm optimization engine, which includes a benchmark artificial intelligence algorithm pool and an extended artificial intelligence algorithm group, comprises: loading predefined benchmark artificial intelligence algorithms from a machine learning library to form a benchmark artificial intelligence algorithm pool; searching a hyperparameter space through an automatic machine learning framework to generate an extended artificial intelligence algorithm group; configuring a parallel execution environment for the benchmark artificial intelligence algorithm pool and the extended artificial intelligence algorithm group to ensure that each algorithm independently processes input data; setting a standardized interface for algorithm output results. 4.The shield tunneling full deformation prediction method based on artificial intelligence according to claim 1, wherein, the error index between the calculation prediction result and the historical true value, comprising: normalizing the prediction result of each data block; extracting the actual deformation amount of the corresponding data block in the historical true value, calculating the absolute error and the relative error; based on time series analysis technology, calculating the dynamic time warping distance between the prediction result and the actual deformation amount; aggregating the absolute error, the relative error and the dynamic time warping distance to form a comprehensive error index. 5.The shield tunneling full deformation prediction method based on artificial intelligence according to claim 1, wherein, the best prediction algorithm is selected by using a trade-off decision mechanism, comprising: multi-dimensional standardization of the adaptive metric set to generate a standardized adaptive matrix; defining a constraint function of real-time construction restriction conditions and applying the constraint function to the standardized adaptive matrix; based on the standardized adaptive matrix, calculating the comprehensive utility value of each algorithm through multi-attribute utility theory, and selecting the algorithm with the highest comprehensive utility value as the best prediction algorithm. 6.The shield tunneling full deformation prediction method based on artificial intelligence according to claim 1, wherein, the full deformation prediction model is generated by using the best prediction algorithm and the dominant feature group, comprising: using the dominant feature group to perform feature selection on the training data to obtain a reduced training set; inputting the reduced training set into the best prediction algorithm for model training, adjusting the hyperparameters to optimize the performance; verifying the prediction accuracy of the trained model on the test set; exporting the model parameters and structure of the training completed model, and packaging it as a full deformation prediction model. 7.The shield tunneling full deformation prediction method based on artificial intelligence according to claim 1, wherein, the full deformation prediction model is integrated into the shield construction monitoring platform, comprising: converting the full deformation prediction model into a lightweight inference module to adapt to the computing resources of the monitoring platform; configuring a model input interface to receive real-time construction data stream, and setting an output interface to send prediction results. 8.The shield tunneling full deformation prediction method based on artificial intelligence of claim 3, wherein, the hyperparameter space is searched through an automatic machine learning framework, comprising: defining the search range of hyperparameters and setting the search step; using the Bayesian optimization algorithm to explore the combination of hyperparameters, recording the performance of each hyperparameter combination on the validation set, and dynamically adjusting the search direction according to the validation results.

Images