Construction parameter optimization method based on large model and meta learning

Abstract

The invention relates to the technical field of TBM construction, in particular to a construction parameter optimization method based on a large model and meta learning. A rock ballast feature data set is extracted through a primary SAM large model, and the model is constructed by using a meta-learning algorithm, so that the method is endowed with strong environmental adaptability, various complex construction environments and frequently changing lithology can be quickly adapted, and the reliability of the construction process is greatly improved; according to the extracted rock ballast feature data set and the identified lithology category features, construction parameters are optimized, the measure is of far-reaching significance, material loss in the construction process can be reduced, the construction risk is reduced, the tunneling efficiency of the TBM can be remarkably improved, stable equipment performance is guaranteed, the service life of the equipment is prolonged, the maintenance cost of the equipment is reduced, and meanwhile, the construction efficiency of the TBM is improved. The construction safety level is improved, a powerful guarantee is provided for smooth development of complex projects such as deep-buried tunnels and river-crossing and sea-crossing tunnels, and the tunnel construction industry is promoted to advance to the development stage of higher quality and higher efficiency.

Description

Technical Field

[0001] The present invention relates to the field of TBM construction technology, and in particular to a construction parameter optimization method based on a large model and meta-learning. Background Art

[0002] With the rapid development of TBM (Tunnel Boring Machine) construction technology, complex projects such as deep tunnels and cross-river and cross-sea tunnels have become widely implemented. However, the construction environment of these major projects is complex, and the rock properties change frequently and in various forms during excavation. This has a direct impact on TBM excavation efficiency, equipment performance, and construction safety.

[0003] Existing rock ballast analysis technology has significant flaws. It primarily relies on subjective judgment of rock ballast morphology, volume, and lithology based on human experience. This is not only inefficient but also prone to misjudgment. Furthermore, current single-sensor data acquisition methods (e.g., relying solely on images or point clouds) cannot fully capture the true shape of rock ballast. Furthermore, complex environments, such as dust, humidity, and light variations, interfere with traditional methods, making it difficult to achieve high-precision analysis.

[0004] Therefore, in modern TBM construction, there is an urgent need for a new method that can integrate multimodal data (such as images and point clouds) and adapt to complex working conditions to achieve real-time analysis to meet the urgent need for rapid and accurate analysis of rock slag characteristics. Especially considering the complex geological formations and their significant impact on construction, rapid learning and discrimination technology is essential for responding to these challenges in a timely and effective manner. Summary of the Invention

[0005] In order to solve the above shortcomings in the prior art, the present invention proposes a construction parameter optimization method based on a large model and meta-learning.

[0006] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0007] A construction parameter optimization method based on a large model and meta-learning comprises: obtaining real-time rock slag data recorded during a real-time construction process; preprocessing the real-time rock slag data to obtain preliminary rock slag data; extracting features from the preliminary rock slag data based on a preset primary SAM large model to obtain a rock slag feature data set; constructing a meta-learning model using a meta-learning algorithm; identifying lithologic categories of the preliminary rock slag data based on the meta-learning model to obtain lithologic category features; obtaining construction parameters of the real-time construction process; and optimizing the construction parameters based on the rock slag feature data set and the lithologic category features to obtain optimized construction parameters. By integrating the SAM large model and the meta-learning framework, the developed method can quickly adapt to various complex construction environments and frequent rock property changes, achieve accurate labeling of rock slag morphology, and solve the problem of rock property identification, greatly improving the degree of automation of data processing and significantly improving the reliability of the construction process. Compared with traditional methods, the meta-learning model constructed using the meta-learning algorithm has excellent robustness, high efficiency and real-time performance. The meta-learning model can "learn to learn" and has the significant advantage of using existing knowledge to quickly acquire new knowledge. It can quickly adapt to various complex construction environments and frequently changing rock properties, greatly improving The reliability of the construction process; the SAM large model has high-precision and universal segmentation capabilities. It optimizes construction parameters based on the rock debris feature dataset it extracts and the rock type category characteristics identified by the meta-learning model. This measure is of far-reaching significance. It can not only reduce material loss and construction risks during the construction process, but also significantly improve the TBM excavation efficiency, ensure stable equipment performance, extend equipment service life, and reduce equipment maintenance costs. At the same time, the construction safety level can be improved, which provides a strong guarantee for the smooth development of complex projects such as deep buried tunnels and cross-river and cross-sea tunnels, and promotes the tunnel construction industry to a higher quality and more efficient development stage.

[0008] Furthermore, the feature extraction of preliminary rock slag data based on the preset primary SAM large model to obtain a rock slag feature data set includes: optimizing the preliminary rock slag data based on the preset multimodal data fusion model to obtain optimized rock slag data; segmenting and labeling the optimized rock slag data based on the primary SAM large model to obtain a labeled data set and prompt information; extracting all segmentation masks in the labeled data set according to the prompt information to obtain a segmentation mask set; calculating the rock slag volume features, rock slag boundary features and rock slag morphology features based on the preset three-dimensional reconstruction algorithm and the segmentation mask set; and generating a rock slag feature data set according to the rock slag volume features, rock slag boundary features and rock slag morphology features. Based on the prompt information, the segmentation mask set is extracted, and combined with the three-dimensional reconstruction algorithm, the volume, boundary, morphology and other characteristics of the rock slag are calculated, comprehensively presenting the physical characteristics of the rock slag, helping construction personnel to fully understand the rock slag situation; the final generated rock slag feature data set provides a key basis for construction decision-making. Construction personnel use this to optimize construction parameters, arrange progress, and formulate response plans, effectively improving construction efficiency, reducing construction risks, and ensuring the smooth progress of complex projects such as deep buried tunnels and cross-river and cross-sea tunnels.

[0009] Furthermore, the lithologic category identification of the preliminary rock slag data according to the meta-learning model to obtain lithologic category characteristics includes: obtaining geological survey reports and historical rock slag data from a preset database to obtain a first large sample set; performing primary internal and external loop training on the meta-learning model according to the first large sample set to obtain an empirical meta-model; lithologic classification of the preliminary rock slag data based on the empirical meta-model to obtain a first classification result; obtaining a confidence level from the first classification result to obtain a first confidence level; determining whether the first confidence level is lower than a preset first threshold; and determining whether the first confidence level is lower than a preset first threshold when the first confidence level is lower than the first threshold. The verification meta-model is optimized in an inner loop to obtain an improved meta-model; the preliminary rock slag data is lithologically classified by the improved meta-model to obtain a second classification result; the confidence level is obtained from the second classification result to obtain a second confidence level; it is determined whether the second confidence level is lower than a preset second threshold; when the second confidence level is not lower than the second threshold, the lithology category characteristics are generated according to the second classification result; when the second confidence level is lower than the second threshold, the improved meta-model is optimized in a secondary inner and outer loop to obtain an optimized meta-model; the preliminary rock slag data is identified in lithology category by the optimized meta-model to obtain the lithology category characteristics. The combination of geological survey reports and historical rock chip data makes the first large sample set cover a wide range and improves the adaptability of the model to different lithologies. Based on the first large sample set, the meta-learning model is trained in primary internal and external loops, which enables the model to learn common features and laws in complex data, obtain an empirical meta-model, and have preliminary lithology classification capabilities. Then, based on the confidence level of the classification results, if the first confidence level is lower than the first threshold, the empirical meta-model is optimized in a secondary internal loop, which can quickly adjust the model parameters for the newly emerged lithology categories and improve the recognition ability of special lithologies. If the second confidence level is lower than the second threshold, Secondary internal and external loop optimization is carried out, and the model is comprehensively updated, which greatly improves the accuracy of model classification and enables it to adapt to various complex and changeable rock properties; the lithology category of the preliminary rock slag data is identified through the optimized meta-model. Whether it is direct identification through the optimized meta-model or the generation of lithology category characteristics based on the second classification results when the second confidence level is not lower than the second threshold, accurate lithology information can be obtained, which provides a key basis for construction decision-making. Construction personnel can reasonably select construction equipment and processes according to the lithology, plan the construction schedule, effectively reduce construction risks, improve construction efficiency, and ensure the smooth progress of the project.

[0010] Furthermore, when the first confidence level is lower than the first threshold, the empirical metamodel is subjected to a secondary inner loop optimization to obtain an improved metamodel, including: when the first confidence level is lower than the first threshold, the preliminary rock chip data is enhanced and labeled according to the data enhancement method to obtain a new lithologic data set; the new lithologic data set is divided according to a preset first lithologic classification task to obtain a first task set; a support set is obtained from the first task set according to a preset first ratio; and the empirical metamodel is subjected to a secondary inner loop optimization according to a preset first gradient threshold and support set to obtain an improved metamodel. The preliminary rock ballast data is enhanced through data augmentation, expanding the amount of new category data, enriching sample diversity, and significantly improving the representation capability of new lithologic category features. This provides higher-quality, more representative data for subsequent model training, effectively compensating for the shortcomings of the original data in the new category. A two-level inner-loop optimization of the empirical metamodel is performed by combining the support set and the first gradient threshold, without performing outer-loop weight updates. This ensures the empirical metamodel's rapid adaptation to new lithologic categories without affecting its generalization performance for existing tasks. This inner-loop "task-level hierarchical optimization" enables the model to efficiently learn from a small amount of task data, laying a solid foundation for continuous adaptation to new tasks. Through a confidence feedback mechanism and online learning, the metamodel can dynamically learn, continuously expand its knowledge base, and accurately identify new lithologic category features. The active learning mechanism avoids interference from a large number of irrelevant samples. In TBM construction environments, this not only significantly improves the accuracy of lithologic category feature recognition but also provides reliable data support for construction parameter optimization, such as thrust and speed adjustment, effectively improving construction efficiency and enhancing construction safety, thus possessing extremely broad application prospects.

[0011] Furthermore, when the second confidence level is lower than the second threshold, the improved metamodel is subjected to secondary internal and external loop optimization to obtain an optimized metamodel, including: when the second confidence level is lower than the second threshold, the new lithologic data set and the first large sample set are merged to obtain a second large sample set; the second large sample set is divided according to a preset second lithologic classification task to obtain a second task set; and the improved metamodel is subjected to secondary internal and external loop optimization according to the second task set to obtain an optimized metamodel. Merging the new lithologic dataset and the first large sample set into the second large sample set achieves a deep fusion of new and old data. The new lithologic dataset enriches the first large sample set's data on new lithologic categories, while the first large sample set provides broader context for the new data. The combination of the two forms a more comprehensive and rich data system, laying a solid foundation for subsequent model optimization. The second large sample set is divided according to the pre-set second lithologic classification task to obtain the second task set, which is then used to perform secondary internal and external loop optimization on the improved meta-model. This deep optimization not only considers the new lithologic data but also incorporates historical data, enabling the model to learn on a more complex dataset and comprehensively adjust parameters, further improving the meta-model's ability to identify various lithologic types and enhancing the model's stability and generalization performance. The optimized meta-model can more accurately identify lithologic category characteristics. During TBM construction, construction parameters, such as thrust and rotation speed, can be more rationally adjusted based on more accurate lithologic information, effectively avoiding construction risks caused by lithologic misjudgment, significantly improving construction efficiency, ensuring construction safety, and providing a strong guarantee for the smooth progress of the project.

[0012] Furthermore, the optimized rock slag data is segmented and labeled based on the primary SAM large model to obtain a labeled data set and prompt information, including: configuring the primary SAM large model with an image encoder, a prompt encoder, a mask generator and a mask decoder to obtain a secondary SAM model; identifying prompt information and multi-scale features in the optimized rock slag data according to the secondary SAM model; and segmenting and labeling the optimized rock slag data according to the prompt information, multi-scale features and the secondary SAM model to obtain a labeled data set. The primary SAM model has excellent generalization capabilities and can adapt to various types of complex image data, providing a solid model foundation for rock slag data processing. Next, the configured secondary SAM model is used to identify and optimize the prompt information and multi-scale features in the rock slag data, greatly improving the recognition accuracy of key areas of rock slag and accurately extracting rich and precise features; finally, based on this information and the secondary SAM model, the optimized rock slag data is segmented and labeled to generate a high-quality labeled data set. This process not only clearly presents the rock slag data structure, but also provides accurate data support for subsequent lithology analysis and construction parameter optimization, helping construction personnel to accurately grasp the rock properties, reasonably adjust construction plans, effectively improve construction efficiency, reduce construction risks, and effectively ensure the smooth progress of the project.

[0013] Furthermore, the rock slag volume features, rock slag boundary features and rock slag morphological features are calculated based on a preset three-dimensional reconstruction algorithm and a segmentation mask set, including: mapping all segmentation masks in the segmentation mask set to a three-dimensional point cloud space to obtain a valid point cloud subset; creating a three-dimensional morphological model based on the three-dimensional reconstruction algorithm and the valid point cloud subset; performing rock slag volume calculation on the annotated data set according to the rock slag volume calculation method and the three-dimensional morphological model to obtain rock slag volume features; performing geometric feature analysis on the rock slag volume features according to the three-dimensional morphological model to obtain rock slag morphological features and rock slag boundary features. The segmentation masks in the segmentation mask set are accurately mapped to the three-dimensional point cloud space to obtain a valid point cloud subset, realizing efficient conversion from two-dimensional to three-dimensional data, exploring the potential value of the mask, and laying a solid foundation for subsequent analysis; then, a three-dimensional morphology model is created based on the three-dimensional reconstruction algorithm and the valid point cloud subset to intuitively present the spatial structure and shape contour of the rock slag, providing a precise visualization carrier for quantitative analysis and feature extraction, and improving the accuracy of rock slag research; then, the model is used to calculate the volume characteristics of the rock slag, and geometric feature analysis is performed to obtain the rock slag morphological characteristics and rock slag boundary characteristics, and comprehensively characterize the physical properties of the rock slag; accurate rock slag volume characteristics, morphological characteristics, and boundary characteristics provide strong support for construction decision-making. Based on these characteristics, construction strategies can be scientifically adjusted, such as reasonably arranging the excavation sequence and accurately controlling the blasting force, effectively avoiding construction delays or safety hazards caused by inaccurate understanding of rock slag characteristics, significantly improving construction efficiency, reducing construction risks, providing reliable guarantees for the smooth progress of TBM construction, and promoting the high-quality completion of the project.

[0014] Furthermore, the improved metamodel is subjected to secondary inner and outer loop optimization according to the second task set to obtain an optimized metamodel, including: obtaining a first training set from the second task set according to a preset second ratio; performing inner loop optimization on the improved metamodel according to the first training set and a preset second gradient threshold to obtain a first task weight set and an adaptive metamodel; obtaining a first verification set from the second task set according to a preset third ratio; performing secondary inner and outer loop optimization on the adaptive metamodel according to the first task weight set and the first verification set to obtain a loss value and an enhanced metamodel; and performing a global weight update on the enhanced metamodel according to the loss value to obtain an optimized metamodel. By performing the ultimate inner loop optimization on the improved meta-model, the weights of specific tasks in the second task set are quickly adjusted so that it can quickly adapt to new data. Then, the ultimate outer loop optimization is performed using the first task weight set and the first validation set to achieve "meta-level optimization." The new task knowledge is integrated into the adaptive meta-model to enhance its adaptability to similar tasks in the future while maintaining its memory of old tasks. The meta-model's generalization ability is comprehensively improved by combining the inner and outer loops to perform local and global weight updates on the meta-model. Ultimately, the optimized meta-model obtained based on the loss value can more accurately and quickly identify lithologic category characteristics, provide construction personnel with a more reliable decision-making basis, and effectively ensure the efficient and safe implementation of TBM construction.

[0015] Furthermore, the meta-learning model is subjected to primary inner and outer loop training according to the first large sample set to obtain an empirical meta-model, including: dividing the first large sample set according to a preset third lithologic classification task to obtain a third task set; obtaining a second training set from the third task set according to a preset fourth ratio; performing primary inner loop training on the meta-learning model according to the second training set and a preset third gradient threshold to obtain a primary meta-model; obtaining a second validation set from the third task set according to a preset fifth ratio; evaluating the primary meta-model according to the second validation set to obtain an evaluation result; and performing primary outer loop training on the primary meta-model according to the third task set and the evaluation result to obtain an empirical meta-model. The first large sample set is divided according to the third lithology classification task to obtain the third task set, which provides a rich and diverse task scenario for model training. The second training set and the second validation set are obtained from them in different proportions, providing targeted data support for model training and evaluation. The meta-learning model is trained in a primary inner loop using the second training set and the third gradient threshold to generate a primary meta-model. In this process, the meta-learning model learns "how to quickly adapt to new tasks" and can accurately classify initial lithology categories. Then, the primary meta-model is trained in a primary outer loop to obtain an empirical meta-model. This training method greatly improves the generalization ability of the primary meta-model, ensuring that when faced with data containing new lithology, the model can quickly complete lithology identification with very few labeled samples. Ultimately, the optimized empirical meta-model provides construction personnel with more accurate lithology identification results, so that construction personnel can adjust construction plans more scientifically, such as reasonably selecting construction tools and determining construction processes, effectively improving construction efficiency, reducing construction risks, and ensuring the safe and smooth progress of TBM construction.

[0016] Furthermore, the preliminary rock slag data is optimized based on a preset multimodal data fusion model to obtain optimized rock slag data, including: fusing the preliminary rock slag data through a multimodal data fusion model to obtain multimodal data; obtaining calibration data of the camera and external parameters of the laser scanner from a real-time construction process; uniformly mapping the multimodal data according to the calibration data and external parameters to obtain aligned data; and optimizing the aligned data using a multimodal data fusion model to obtain optimized rock slag data. The preliminary rock slag data is fused using a multimodal data fusion model and input into a deep learning model to integrate multiple data information and improve data richness; camera calibration data and laser scanner external parameters are obtained from real-time construction to provide key parameters for subsequent data processing; based on these data, the multimodal data is uniformly mapped to generate aligned data, giving the rock slag data depth information to make the data more three-dimensional and accurate; finally, the multimodal data fusion model is used again to optimize the aligned data to obtain optimized rock slag data, further remove noise, fill in gaps, and improve data quality; high-quality optimized rock slag data provides a solid basis for subsequent rock property identification, construction decision-making, etc., helping construction personnel to accurately grasp the construction situation, reasonably plan the construction process, improve construction efficiency, reduce construction risks, and ensure the smooth progress of the project.

[0017] The beneficial effects of the construction parameter optimization method based on a large model and meta-learning of the present invention are as follows:

[0018] By integrating the SAM large model and the meta-learning framework, the developed method can quickly adapt to various complex construction environments and frequent rock property changes, achieve accurate labeling of rock slag morphology, and solve the problem of rock property identification, greatly improving the degree of automation of data processing and significantly improving the reliability of the construction process. Compared with traditional methods, the meta-learning model constructed using the meta-learning algorithm has excellent robustness, high efficiency and real-time performance. The meta-learning model can "learn to learn" and has the significant advantage of using existing knowledge to quickly acquire new knowledge. It can quickly adapt to various complex construction environments and frequently changing rock properties, greatly improving The reliability of the construction process; the SAM large model has high-precision and universal segmentation capabilities. It optimizes construction parameters based on the rock debris feature dataset it extracts and the rock type category characteristics identified by the meta-learning model. This measure is of far-reaching significance. It can not only reduce material loss and construction risks during the construction process, but also significantly improve the TBM excavation efficiency, ensure stable equipment performance, extend equipment service life, and reduce equipment maintenance costs. At the same time, the construction safety level can be improved, which provides a strong guarantee for the smooth development of complex projects such as deep buried tunnels and cross-river and cross-sea tunnels, and promotes the tunnel construction industry to a higher quality and more efficient development stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0020] Figure 1 A first flow chart of a construction parameter optimization method based on a large model and meta-learning provided by an embodiment of the present invention;

[0021] Figure 2 A second flow chart of a construction parameter optimization method based on a large model and meta-learning provided by an embodiment of the present invention;

[0022] Figure 3 A third flow chart of a construction parameter optimization method based on a large model and meta-learning provided by an embodiment of the present invention;

[0023] Figure 4 A fourth flow chart of a construction parameter optimization method based on a large model and meta-learning provided by an embodiment of the present invention;

[0024] Figure 5 A fifth flow chart of a construction parameter optimization method based on a large model and meta-learning provided by an embodiment of the present invention;

[0025] Figure 6 A sixth flow chart of a construction parameter optimization method based on a large model and meta-learning provided by an embodiment of the present invention;

[0026] Figure 7 A seventh flow chart of a construction parameter optimization method based on a large model and meta-learning provided by an embodiment of the present invention;

[0027] Figure 8 An eighth flow chart of a construction parameter optimization method based on a large model and meta-learning provided by an embodiment of the present invention;

[0028] Figure 9 An eighth flow chart of a construction parameter optimization method based on a large model and meta-learning provided by an embodiment of the present invention;

[0029] Figure 10 A structural schematic diagram of a fracture zone stratum identification device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0031] The terms "first," "second," "third," "fourth," and the like (if any) in the description and claims of the present invention and in the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatus.

[0032] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In an embodiment of the present invention, a construction parameter optimization method based on a large model and meta-learning includes:

[0033] 101. Obtain real-time rock slag data recorded during the construction process;

[0034] In this embodiment, the real-time rock slag data recorded during the real-time construction process are rock slag point cloud data and rock slag image data collected by a laser scanner and a high-resolution camera installed on the TBM belt conveyor. The integrated multimodal data (rock slag point cloud data and rock slag image data) can fully present the true form of the rock slag compared to the data acquisition method of a single sensor.

[0035] 102. Preprocessing the real-time rock slag data to obtain preliminary rock slag data;

[0036] In this embodiment, the preprocessing of real-time rock slag data has different processing methods for different types of data; for rock slag point cloud data, denoising, segmentation and calibration operations are mainly performed; while for rock slag image data, denoising and enhancement processing are focused on. In addition, in order to improve the sensitivity of the model to local features, the rock slag image data can be multi-scale segmented, and large rock slag images can be divided into smaller areas, so as to facilitate the subsequent model to capture the local texture features of the rock more finely; the frequency of rock slag point cloud data and rock slag image data acquisition needs to match the conveyor belt speed to ensure data integrity.

[0037] 103. Extract features from the preliminary rock slag data based on the preset primary SAM large model to obtain a rock slag feature data set;

[0038] 104. A meta-learning model is constructed using a meta-learning algorithm;

[0039] In this embodiment, the meta-learning model can "learn to learn" and has the significant advantage of rapidly acquiring new knowledge using existing knowledge. For example, after a meta-learning model has learned the lithology of granite and tuff based on thousands of images, it only needs to identify two or five limestone images to learn the characteristics of limestone. In complex geological backgrounds with frequently changing lithology, it can quickly grasp the lithology category characteristics in the image based on only a small number of limestone images. This can reduce material waste, lower construction risks, significantly improve TBM excavation efficiency, ensure stable equipment performance, extend equipment service life, and reduce equipment maintenance costs. The meta-learning model constructed using the meta-learning algorithm has excellent robustness, efficiency, and real-time performance, providing strong support for the smooth progress of complex projects such as deep tunnels and cross-river and cross-sea tunnels, and promoting the tunnel construction industry to a higher quality and more efficient development stage.

[0040] 105. Perform lithology classification on the preliminary rock slag data based on the meta-learning model to obtain lithology classification characteristics;

[0041] In this embodiment, lithologic classification features include topographic features (such as length, width, area, texture, and roughness index). The meta-learning model has online learning capabilities and is seamlessly integrated with the control system of the construction equipment. During deployment, the initial model is responsible for processing common lithologic classification features. When new lithologic classification features are encountered, the meta-model is updated. The updated meta-model can be deployed to the equipment on the construction site via OTA (over-the-air) updates. This update method enables continuous optimization of the meta-model. In this way, the meta-model will gradually expand its knowledge scope during construction, becoming increasingly adapted to the lithologic distribution characteristics of a specific construction area.

[0042] Feature extraction of preliminary rock ballast data based on the primary SAM large model and the construction of a meta-learning model using a meta-learning algorithm. These advanced technologies enable the method to quickly adapt to different construction environments and rock property changes, improving the reliability of the entire construction process.

[0043] 106. Obtain construction parameters of the real-time construction process;

[0044] In this embodiment, the construction parameters include excavation parameters, such as cutterhead parameters, thrust parameters, rotation speed, and propulsion speed;

[0045] 107. Optimize the construction parameters according to the rock slag characteristic data set and lithology category characteristics to obtain the optimized construction parameters.

[0046] In this embodiment, by optimizing the construction parameters based on the rock ballast characteristic dataset and lithologic category characteristics, more reasonable construction parameters can be obtained, i.e., optimized construction parameters. Optimizing construction parameters can further assist in construction decision-making, which helps reduce losses during construction, lowers construction risks, improves TBM excavation efficiency, ensures equipment performance, and extends equipment service life. It also enhances construction safety, which is of great significance to the smooth implementation of complex projects such as deep tunnels and cross-river and cross-sea tunnels.

[0047] In this embodiment, by integrating the SAM large model and the meta-learning framework, the developed method can quickly adapt to various complex construction environments and frequent rock property changes, achieve accurate labeling of rock slag morphology, and solve the problem of rock property identification, greatly improving the degree of automation of data processing and significantly improving the reliability of the construction process; compared with traditional methods, the meta-learning model constructed using the meta-learning algorithm has excellent robustness, high efficiency and real-time performance. The meta-learning model can achieve "learning to learn" and has the significant advantage of using existing knowledge to quickly acquire new knowledge. It can quickly adapt to various complex construction environments and frequently changing rock properties, greatly The SAM large model has high-precision and universal segmentation capabilities. It optimizes construction parameters based on the rock debris feature dataset extracted and the lithology category characteristics identified by the meta-learning model. This initiative is of far-reaching significance. It not only reduces material loss and construction risks during the construction process, but also significantly improves the excavation efficiency of TBMs, ensures stable equipment performance, extends equipment service life, and reduces equipment maintenance costs. At the same time, the construction safety level is improved, providing a strong guarantee for the smooth implementation of complex projects such as deep buried tunnels and cross-river and cross-sea tunnels, and promoting the tunnel construction industry to a higher quality and more efficient development stage.

[0048] See also Figure 2 A second embodiment of a construction parameter optimization method based on a large model and meta-learning in an embodiment of the present invention includes:

[0049] 201. Optimizing the preliminary rock slag data based on a preset multimodal data fusion model to obtain optimized rock slag data;

[0050] In this embodiment, the multimodal data fusion model is a deep learning model, which can be BiFPN and UNet;

[0051] Optimizing preliminary rock slag data through deep learning models can effectively remove noise and redundant information from the data, fill in missing data, improve data accuracy and completeness, provide higher-quality basic data for subsequent analysis and processing, and greatly enhance the reliability and stability of the entire data processing process.

[0052] 202. Segment and label the optimized rock slag data based on the primary SAM large model to obtain a labeled data set and prompt information;

[0053] 203. Extract all segmentation masks in the labeled data set according to the prompt information to obtain a segmentation mask set;

[0054] 204. Calculate the rock slag volume features, rock slag boundary features, and rock slag morphology features based on a preset three-dimensional reconstruction algorithm and a segmentation mask set;

[0055] 205. Generate a rock slag feature data set based on rock slag volume features, rock slag boundary features and rock slag morphology features.

[0056] In this embodiment, deep learning models such as BiFPN or UNet are used to optimize preliminary rock slag data, remove noise and redundancy, fill in gaps, greatly improve data quality, and lay a solid foundation for subsequent processes; the optimized data is segmented and labeled through the primary SAM large model, accurately identifying different rock slag areas, obtaining rich labeled data and prompt information, and laying the foundation for accurate analysis of rock slag characteristics.

[0057] Based on the prompt information, the segmentation mask set is extracted, and combined with the three-dimensional reconstruction algorithm, the volume, boundary, morphology and other characteristics of the rock slag are calculated, comprehensively presenting the physical characteristics of the rock slag, helping construction personnel to fully understand the rock slag situation; the final generated rock slag feature data set provides a key basis for construction decision-making. Construction personnel use this to optimize construction parameters, arrange progress, and formulate response plans, effectively improving construction efficiency, reducing construction risks, and ensuring the smooth progress of complex projects such as deep buried tunnels and cross-river and cross-sea tunnels.

[0058] See also Figure 3 A third embodiment of a construction parameter optimization method based on a large model and meta-learning in an embodiment of the present invention includes:

[0059] 301. Obtaining geological survey reports and historical rock slag data from a preset database to obtain a first large sample set;

[0060] In this embodiment, the first large sample set is a dataset that has been lithologically annotated. The geological survey report is derived from the annotated data of standard rock samples taken by the geological laboratory, which can reflect the geological conditions. The historical rock slag data contains various lithological characteristics during past construction. The combination of the two makes the first large sample set cover a wide range and improves the adaptability of the model to different lithologies.

[0061] 302. Performing primary inner and outer loop training on the meta-learning model based on the first large sample set to obtain an empirical meta-model;

[0062] 303. Performing lithologic classification on the preliminary rock ballast data based on the empirical metamodel to obtain a first classification result;

[0063] 304. Obtain a confidence level from the first classification result to obtain a first confidence level;

[0064] 305. Determine whether the first confidence level is lower than a preset first threshold;

[0065] 306. When the first confidence level is lower than the first threshold, performing inner loop optimization on the empirical metamodel to obtain an improved metamodel;

[0066] In this embodiment, the first classification result includes a lithologic category feature and a confidence level. When the confidence level is not lower than a first threshold, for example, not lower than 50%, the lithologic category feature can be directly obtained from the first classification result. When the confidence level is lower than the first threshold, it means that a new lithologic category feature has emerged during the lithologic classification process, and the empirical metamodel needs to be quickly updated to adapt to the new lithologic category feature.

[0067] 307. Perform lithologic classification on the preliminary rock ballast data by improving the meta-model to obtain a second classification result;

[0068] 308. Obtaining a confidence level from the second classification result to obtain a second confidence level;

[0069] 309. Determine whether the second confidence level is lower than a preset second threshold;

[0070] 310. When the second confidence level is not lower than the second threshold, generating a lithology category feature according to the second classification result;

[0071] In this embodiment, when the second confidence level is not lower than the second threshold, the meta-model may only execute the secondary inner loop for optimization and updating to quickly adapt to the recognition of lithologic category features in the new task;

[0072] 311. When the second confidence level is lower than the second threshold, performing secondary inner and outer loop optimization on the improved metamodel to obtain an optimized metamodel;

[0073] 312. The lithology category of the preliminary rock slag data is identified by optimizing the meta-model to obtain the lithology category characteristics.

[0074] In this embodiment, after each model update, its confidence level is also increased along with the threshold to reduce unnecessary manual intervention. The meta-model MAML usually uses a pre-trained deep convolutional neural network model as its feature extractor, such as ResNet or EfficientNet. These deep convolutional neural network models are pre-trained on natural images and can well capture features such as color, shape, and texture. By inputting rock slag image data into these deep convolutional neural network models, their high-level feature vectors are extracted while retaining detailed texture information. By configuring the deep convolutional neural network model with an attention mechanism (such as an SE module or a CBAM module), the deep convolutional neural network model can enhance its attention to key areas, such as the edge features of rock slag or areas with obvious texture differences. The meta-model MAML adjusts the parameters of these feature extractors to enable the meta-model MAML to quickly adapt to new classification tasks. During the training phase of the meta-model MAML, the parameters of the deep convolutional neural network model are adjusted through meta-optimization. The meta-model MAML optimizes the initial values ​​of the deep convolutional neural network model parameters so that it can perform well on new tasks after a small amount of training data and a few gradient updates.

[0075] The combination of geological survey reports and historical rock chip data makes the first large sample set cover a wide range and improves the adaptability of the model to different lithologies. Based on the first large sample set, the meta-learning model is trained in primary internal and external loops, which enables the model to learn common features and laws in complex data, obtain an empirical meta-model, and have preliminary lithology classification capabilities. Then, based on the confidence level of the classification results, if the first confidence level is lower than the first threshold, the empirical meta-model is optimized in a secondary internal loop, which can quickly adjust the model parameters for the newly emerged lithology categories and improve the recognition ability of special lithologies. If the second confidence level is lower than the second threshold, Secondary internal and external loop optimization is carried out, and the model is comprehensively updated, which greatly improves the accuracy of model classification and enables it to adapt to various complex and changeable rock properties; the lithology category of the preliminary rock slag data is identified through the optimized meta-model. Whether it is direct identification through the optimized meta-model or the generation of lithology category characteristics based on the second classification results when the second confidence level is not lower than the second threshold, accurate lithology information can be obtained, which provides a key basis for construction decision-making. Construction personnel can reasonably select construction equipment and processes according to the lithology, plan the construction schedule, effectively reduce construction risks, improve construction efficiency, and ensure the smooth progress of the project.

[0076] See also Figure 4 A fourth embodiment of a construction parameter optimization method based on a large model and meta-learning in the embodiments of the present invention includes:

[0077] 401. When the first confidence level is lower than a first threshold, performing enhanced annotation on the preliminary rock slag data according to a data enhancement method to obtain a new lithologic data set;

[0078] In this embodiment, data enhancement methods are used to enhance the data with low recognition accuracy in the preliminary rock ballast data, expand the amount of data of the new category, and increase the diversity of the new category samples. For example, data enhancement methods such as adjusting image brightness, cropping, and rotation are used to enhance the representation ability of the new lithologic category features. The preliminary rock ballast data are then labeled as new lithologic data to obtain a new lithologic dataset.

[0079] 402. Divide the new lithology dataset according to a preset first lithology classification task to obtain a first task set;

[0080] In this embodiment, the first lithology classification task involves systematically partitioning the new lithology dataset based on specific rules and objectives. It categorizes the rock debris samples in the new lithology dataset into different subsets based on differences in lithology properties such as mineral composition, structural characteristics, and hardness. Each subset corresponds to a specific lithology category.

[0081] 403. Obtain a support set from the first task set according to a preset first ratio;

[0082] 404. Perform secondary inner loop optimization on the empirical meta-model according to the preset first gradient threshold and support set to obtain an improved meta-model.

[0083] In this embodiment, the empirical meta-model is efficiently trained with a small amount of labeled data (support set), which significantly reduces the cost of data collection. Only a few gradient updates are needed to complete the training of the empirical meta-model for a new category of data sets. The empirical meta-model can also quickly update its weights through a secondary inner loop. During this process, the outer loop weights of the meta-model are not updated, so the generalization performance of the meta-model for old tasks is not affected. The weight update of the inner loop is the "task-level hierarchical optimization" in the meta-learning framework. The empirical meta-model with updated weights can quickly adapt to the new lithology category recognition in the new task in the next task. For new lithology categories, the secondary inner loop can also be used to update the weights of the meta-model. Loop optimization rapidly adjusts the model's parameters to adapt it to new data. The purpose of inner-loop optimization is to enable the meta-model to quickly learn from a small amount of task data, thereby adapting to new tasks. Through a confidence feedback mechanism and an online learning mechanism, the meta-model can continuously expand its knowledge base through dynamic learning to adapt to the identification of new lithologic category features. Furthermore, the meta-model's active learning mechanism avoids interference from a large number of irrelevant samples. This approach has broad application prospects in TBM construction environments. It can not only improve the accuracy of lithologic category feature identification, but also provide reliable data support for the optimization of construction parameters (such as thrust and speed adjustment), thereby improving construction efficiency and safety.

[0084] In this embodiment, preliminary rock ballast data is enhanced using a data augmentation method. This expands the amount of new category data, enriches sample diversity, and significantly improves the representation capability of new lithologic category features, providing higher-quality and more representative data for subsequent model training. This effectively compensates for the shortcomings of the original data in the new category. A two-level inner loop optimization is performed on the empirical metamodel using the support set and the first gradient threshold, without performing an outer loop weight update. This ensures the rapid adaptation of the empirical metamodel to new lithologic categories without affecting its generalization performance for existing tasks. This inner loop "task-level hierarchical optimization" enables the model to efficiently learn from a small amount of task data, laying a solid foundation for continuous adaptation to new tasks. Through a confidence feedback mechanism and an online learning mechanism, the metamodel can dynamically learn, continuously expand its knowledge base, and accurately identify new lithologic category features. The active learning mechanism avoids interference from a large number of irrelevant samples. In a TBM construction environment, this not only significantly improves the accuracy of lithologic category feature recognition, but also provides reliable data support for construction parameter optimization, such as thrust and speed adjustment, thereby effectively improving construction efficiency and enhancing construction safety. It has extremely broad application prospects.

[0085] See also Figure 5 A fifth embodiment of a construction parameter optimization method based on a large model and meta-learning in an embodiment of the present invention includes:

[0086] 501. When the second confidence level is lower than a second threshold, merging the new lithologic dataset and the first large sample set to obtain a second large sample set;

[0087] 502. Divide the second largest sample set according to the preset second lithologic classification task to obtain a second task set;

[0088] 503. Perform secondary inner and outer loop optimization on the improved meta-model according to the second task set to obtain an optimized meta-model.

[0089] In this embodiment, the new lithologic dataset and the first large sample set are merged into the second large sample set, achieving a deep fusion of new and old data. The new lithologic dataset enriches the first large sample set's data on the new lithologic category, while the first large sample set provides broader contextual support for the new data. The combination of the two forms a more comprehensive and rich data system, laying a solid data foundation for subsequent model optimization. The second large sample set is divided according to the preset second lithologic classification task to obtain the second task set, which is then used to perform secondary internal and external loop optimization on the improved meta-model. This deep optimization not only considers the new lithologic data but also incorporates historical data, enabling the model to learn on more complex datasets and comprehensively adjust parameters, further improving the meta-model's ability to identify various lithologic types and enhancing the model's stability and generalization performance. The optimized meta-model can more accurately identify lithologic category characteristics. In TBM construction, construction parameters, such as thrust and rotation speed, can be more rationally adjusted based on more accurate lithologic information, effectively avoiding construction risks caused by lithologic misjudgment, significantly improving construction efficiency, ensuring construction safety, and providing a strong guarantee for the smooth progress of the project.

[0090] See also Figure 6 A sixth embodiment of a construction parameter optimization method based on a large model and meta-learning in the embodiments of the present invention includes:

[0091] 601. Configuring an image encoder, a hint encoder, a mask generator, and a mask decoder for the primary SAM large model to obtain a secondary SAM model;

[0092] 602. Identify and optimize prompt information and multi-scale features in rock ballast data based on the secondary SAM model;

[0093] 603. Segment and label the optimized rock slag data according to the prompt information, multi-scale features, and the secondary SAM model to obtain a labeled data set.

[0094] In this embodiment, the prompt information includes points, boxes, text, etc. Integrating this prompt information into the segmentation and annotation process can improve the secondary SAM model's ability to accurately identify and locate the region of interest. The mask decoder integrates the multi-scale features from the encoder, and the mask generator combines the prompt information to generate a high-quality segmentation mask to complete the segmentation and annotation of the data in the area; based on the fusion of multi-scale features, the mask generator can refine the boundary features in the multi-scale features to avoid over-segmentation or under-segmentation problems; in this process, the secondary SAM model is pre-trained on billions of images and segmentation mask pairs. These data cover various scenes, target shapes and sizes, giving the secondary SAM model excellent generalization capabilities.

[0095] Furthermore, the Level 2 SAM model is not limited to segmentation of certain specific categories. It is a universal segmentation model that can perform high-quality segmentation on objects in almost all scenarios. It performs consistently across images of varying resolutions and maintains excellent performance on unseen data distributions or image types. It can simultaneously segment multiple objects in an image and accurately distinguish complex overlapping regions between objects. These advantages of the Level 2 SAM model are attributed to its pre-training coverage of diverse scenarios. The structural design of the Level 2 SAM model greatly enhances its generalization capabilities across different tasks and datasets.

[0096] In this embodiment, the primary SAM model has excellent generalization capabilities and can adapt to various types of complex image data, providing a solid model foundation for rock slag data processing. Next, the configured secondary SAM model is used to identify and optimize the prompt information and multi-scale features in the rock slag data, greatly improving the recognition accuracy of key areas of rock slag and accurately extracting rich and precise features; finally, based on this information and the secondary SAM model, the optimized rock slag data is segmented and labeled to generate a high-quality labeled data set. This process not only clearly presents the rock slag data structure, but also provides accurate data support for subsequent lithology analysis and construction parameter optimization, helping construction personnel to accurately grasp the rock properties, reasonably adjust the construction plan, effectively improve construction efficiency, reduce construction risks, and effectively ensure the smooth progress of the project.

[0097] See also Figure 7 A seventh embodiment of a construction parameter optimization method based on a large model and meta-learning in the embodiments of the present invention includes:

[0098] 701. Map all segmentation masks in the segmentation mask set to the three-dimensional point cloud space to obtain a valid point cloud subset;

[0099] In this embodiment, accurate conversion from two-dimensional mask information to three-dimensional spatial data is achieved. This step provides a direct and effective data foundation for subsequent three-dimensional space-based analysis, deeply explores the potential value of the segmentation mask, and enables the data to be presented in a more intuitive and three-dimensional form.

[0100] 702. Create a 3D shape model based on a 3D reconstruction algorithm and a valid point cloud subset;

[0101] In this embodiment, the three-dimensional morphology model comprehensively displays the spatial structure, shape and contour of the rock ballast, providing a visual and accurate carrier for subsequent quantitative analysis and feature extraction, greatly improving the accuracy and scientific nature of rock ballast morphology research.

[0102] 703. Calculate the rock slag volume on the labeled data set according to the rock slag volume calculation method and the three-dimensional morphology model to obtain rock slag volume characteristics;

[0103] 704. Perform geometric feature analysis on the volume features of the rock slag based on the three-dimensional morphological model to obtain the rock slag morphological features and rock slag boundary features.

[0104] In this embodiment, the segmentation masks in the segmentation mask set are precisely mapped to the three-dimensional point cloud space to obtain a valid point cloud subset, achieving efficient conversion from two-dimensional to three-dimensional data, exploring the potential value of the mask, and laying a solid foundation for subsequent analysis. Next, a three-dimensional morphology model is created based on the three-dimensional reconstruction algorithm and the valid point cloud subset, intuitively presenting the spatial structure and shape contours of the rock slag, providing a precise visualization carrier for quantitative analysis and feature extraction, and improving the accuracy of rock slag research. The model is then used to calculate the volume characteristics of the rock slag and perform geometric feature analysis to obtain the rock slag morphological characteristics and rock slag boundary characteristics, comprehensively characterizing the rock slag physical properties. Accurate rock slag volume characteristics, morphological characteristics, and boundary characteristics provide strong support for construction decision-making. Based on these characteristics, construction strategies can be scientifically adjusted, such as rationally arranging the excavation sequence and accurately controlling the blasting force, effectively avoiding construction delays or safety hazards caused by inaccurate understanding of rock slag characteristics, significantly improving construction efficiency, reducing construction risks, providing reliable guarantees for the smooth implementation of TBM construction, and promoting the high-quality completion of the project.

[0105] See also Figure 8 An eighth embodiment of a construction parameter optimization method based on a large model and meta-learning in the embodiments of the present invention includes:

[0106] 801. Obtain a first training set from a second task set according to a preset second ratio;

[0107] 802. Perform inner loop optimization on the improved meta-model according to the first training set and a preset second gradient threshold to obtain a first task weight set and an adaptive meta-model;

[0108] In this embodiment, the improved meta-model performs a final inner loop through gradient descent to optimize the specific weights of all tasks in the second task set;

[0109] 803. Obtain a first verification set from the second task set according to a preset third ratio;

[0110] 804. Perform secondary inner and outer loop optimization on the adaptive meta-model according to the first task weight set and the first validation set to obtain a loss value and an enhanced meta-model;

[0111] In this embodiment, the weight update of the outer loop is the "meta-level optimization" in the meta-learning framework, which adjusts the weight of the global adaptive meta-model by integrating task-level information, that is, all task levels; the secondary inner and outer loop optimization is to optimize the adaptive meta-model through the validation sets of multiple tasks. The secondary inner and outer loop optimization can integrate the knowledge of the new task into the adaptive meta-model to obtain an enhanced meta-model, making it better adapted to similar tasks in the future and quickly identifying lithologic category characteristics; the final outer loop can jointly train the adaptive meta-model through multiple tasks to obtain an enhanced meta-model, which includes tasks in the original sample set, so that the enhanced meta-model can retain the ability to remember old tasks;

[0112] 805. Update the global weight of the enhanced meta-model according to the loss value to obtain an optimized meta-model.

[0113] In this embodiment, in order to simultaneously achieve rapid updating and global adaptability of the improved metamodel, the inner loop and outer loop can be combined to perform local and global weight updates on the improved metamodel; when a new lithology category is discovered, the weight of the improved metamodel is first quickly and locally adjusted through the inner loop so that it can immediately adapt to the new data; then the new task is added to the ultimate outer loop optimization to integrate the new knowledge into the metamodel, thereby improving the global generalization ability of the metamodel.

[0114] In this embodiment, the improved meta-model is optimized through an ultimate inner loop, and its weights for specific tasks in the second task set are quickly adjusted so that it can quickly adapt to new data. Then, the first task weight set and the first validation set are used for an ultimate outer loop optimization to achieve "meta-level optimization." The new task knowledge is integrated into the adaptive meta-model to enhance its adaptability to similar tasks in the future while maintaining its memory of old tasks. The meta-model is updated locally and globally through a combination of inner and outer loops, thereby comprehensively improving its generalization ability. Finally, the optimized meta-model obtained based on the loss value can more accurately and quickly identify lithologic category characteristics, provide construction personnel with a more reliable basis for decision-making, and effectively ensure the efficient and safe implementation of TBM construction.

[0115] See also Figure 9 A ninth embodiment of a construction parameter optimization method based on a large model and meta-learning in an embodiment of the present invention includes:

[0116] 901. Divide the first large sample set according to the preset third lithologic classification task to obtain a third task set;

[0117] 902. Obtain a second training set from the third task set according to a preset fourth ratio;

[0118] 903. Performing primary inner loop training on the meta-learning model according to the second training set and a preset third gradient threshold to obtain a primary meta-model;

[0119] In this embodiment, the third task set includes multiple tasks, each task corresponds to a small number of labeled images of a certain lithology. The meta-learning model is trained on different tasks so that the meta-learning model gradually learns "how to quickly adapt to new tasks." For example, in each round of training, a task is randomly selected from the task set, and a small number of samples of the task are used as the training set. The model is trained in an inner loop using the training set, so that the meta-learning model can learn the features of the training set. The trained primary meta-model can accurately classify the initial lithology categories (such as sandstone, mudstone, granite, etc.).

[0120] In another embodiment, in order to enhance the model's ability to capture RGB image details, the meta-learning model can be pre-trained in combination with contrastive learning (method). Contrastive learning generates similar sample pairs by performing different enhancement operations on the image (such as rotation, cropping, color perturbation, etc.). The meta-learning model is then optimized based on the similar sample pairs so that similar sample pairs are brought closer together and different samples are pushed further apart in the feature space. This method can help the meta-learning model learn more discriminative image features. Even when the number of lithology categories is small, the model can effectively distinguish different types of rock debris.

[0121] 904. Obtain a second verification set from the third task set according to a preset fifth ratio;

[0122] 905. Evaluate the primary meta-model according to the second validation set to obtain an evaluation result;

[0123] 906. Perform primary outer loop training on the primary meta-model according to the third task set and the evaluation result to obtain an empirical meta-model.

[0124] In this embodiment, the generalization ability of the primary metamodel can be improved by performing primary outer loop training on the primary metamodel. This training method ensures that when the model is faced with new lithology data, it can quickly complete lithology identification with very few labeled samples.

[0125] The first large sample set is divided according to the third lithology classification task to obtain the third task set, which provides a rich and diverse task scenario for model training. The second training set and the second validation set are obtained from them in different proportions, providing targeted data support for model training and evaluation. The meta-learning model is trained in a primary inner loop using the second training set and the third gradient threshold to generate a primary meta-model. In this process, the meta-learning model learns "how to quickly adapt to new tasks" and can accurately classify initial lithology categories. Then, the primary meta-model is trained in a primary outer loop to obtain an empirical meta-model. This training method greatly improves the generalization ability of the primary meta-model, ensuring that when faced with data containing new lithology, the model can quickly complete lithology identification with very few labeled samples. Ultimately, the optimized empirical meta-model provides construction personnel with more accurate lithology identification results, so that construction personnel can adjust construction plans more scientifically, such as reasonably selecting construction tools and determining construction processes, effectively improving construction efficiency, reducing construction risks, and ensuring the safe and smooth progress of TBM construction.

[0126] See also Figure 10 A tenth embodiment of a construction parameter optimization method based on a large model and meta-learning in an embodiment of the present invention includes:

[0127] 1001: performing data fusion on the preliminary rock ballast data through a multimodal data fusion model to obtain multimodal data;

[0128] In this embodiment, by inputting preliminary rock ballast data into the deep learning model,

[0129] 1002: Obtain camera calibration data and laser scanner external parameters from the real-time construction process;

[0130] 1003: uniformly mapping the multimodal data according to the calibration data and external parameters to obtain aligned data;

[0131] In this embodiment, the point cloud coordinate system and the image coordinate system are uniformly mapped using the camera calibration data and the laser scanner's external parameters. In this process, multimodal data is deeply fused and mapped to be converted into aligned data with depth information.

[0132] 1004: Optimize the aligned data using a multimodal data fusion model to obtain optimized rock slag data.

[0133] The preliminary rock slag data is fused using a multimodal data fusion model and input into a deep learning model to integrate multiple data information and improve data richness; camera calibration data and laser scanner external parameters are obtained from real-time construction to provide key parameters for subsequent data processing; based on these data, the multimodal data is uniformly mapped to generate aligned data, giving the rock slag data depth information to make the data more three-dimensional and accurate; finally, the multimodal data fusion model is used again to optimize the aligned data to obtain optimized rock slag data, further remove noise, fill in gaps, and improve data quality; high-quality optimized rock slag data provides a solid basis for subsequent rock property identification, construction decision-making, etc., helping construction personnel to accurately grasp the construction situation, reasonably plan the construction process, improve construction efficiency, reduce construction risks, and ensure the smooth progress of the project.

[0134] The present invention and its embodiments are described above. This description is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual content is not limited to this. In short, if ordinary technicians in this field are inspired by it and do not depart from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. A construction parameter optimization method based on a large model and meta-learning, characterized in that: include: Obtain real-time rock ballast data recorded during real-time construction; Preprocessing the real-time rock ballast data to obtain preliminary rock ballast data; Based on the preset primary SAM large model, the preliminary rock ballast data is subjected to feature extraction to obtain the rock ballast feature data set; A meta-learning model is constructed using a meta-learning algorithm; The lithology category of the preliminary rock ballast data is identified based on the meta-learning model to obtain the lithology category characteristics; Obtain construction parameters of the real-time construction process; The construction parameters are optimized according to the rock ballast characteristic data set and lithology category characteristics to obtain the optimized construction parameters.

2. A construction parameter optimization method based on a large model and meta-learning as claimed in claim 1, characterized in that: The feature extraction of the preliminary rock slag data based on the preset primary SAM large model to obtain a rock slag feature data set includes: Optimize the preliminary rock slag data based on the preset multimodal data fusion model to obtain optimized rock slag data; Based on the primary SAM large model, the optimized rock slag data is segmented and labeled to obtain the labeled data set and prompt information; Extract all segmentation masks in the labeled dataset according to the prompt information to obtain a segmentation mask set; Based on the preset 3D reconstruction algorithm and segmentation mask set, the rock ballast volume features, rock ballast boundary features and rock ballast morphology features are calculated; A rock slag feature dataset is generated based on the rock slag volume features, rock slag boundary features and rock slag morphology features.

3. The construction parameter optimization method based on large model and meta-learning according to claim 1, characterized in that: The lithology classification of the preliminary rock ballast data is performed according to the meta-learning model to obtain lithology classification characteristics, including: Obtain geological survey reports and historical rock slag data from a pre-set database to obtain the first large sample set; Performing primary inner and outer loop training on the meta-learning model based on the first large sample set to obtain an empirical meta-model; The lithology of the preliminary rock ballast data is classified based on the empirical metamodel to obtain the first classification result; Obtaining a confidence level from the first classification result to obtain a first confidence level; Determining whether the first confidence level is lower than a preset first threshold; When the first confidence level is lower than a first threshold, performing inner loop optimization on the empirical metamodel to obtain an improved metamodel; The lithology classification of the preliminary rock ballast data is carried out by improving the meta-model to obtain the second classification result; Obtaining a confidence level from the second classification result to obtain a second confidence level; Determining whether the second confidence level is lower than a preset second threshold; When the second confidence level is not lower than the second threshold, generating a lithology category feature according to the second classification result; When the second confidence level is lower than a second threshold, performing secondary inner and outer loop optimization on the improved metamodel to obtain an optimized metamodel; The lithology classification of preliminary rock ballast data is identified by optimizing the meta-model to obtain the lithology classification characteristics.

4. A construction parameter optimization method based on a large model and meta-learning as claimed in claim 3, characterized in that: When the first confidence level is lower than the first threshold, performing a secondary inner loop optimization on the empirical metamodel to obtain an improved metamodel, including: When the first confidence level is lower than the first threshold, the preliminary rock ballast data are enhanced and labeled according to a data enhancement method to obtain a new lithology data set; Dividing the new lithology dataset according to a preset first lithology classification task to obtain a first task set; Obtaining a support set from the first task set according to a preset first ratio; The empirical metamodel is optimized on a secondary inner loop according to the preset first gradient threshold and support set to obtain an improved metamodel.

5. A construction parameter optimization method based on a large model and meta-learning as claimed in claim 4, characterized in that: When the second confidence level is lower than the second threshold, performing secondary inner and outer loop optimization on the improved meta-model to obtain an optimized meta-model, including: When the second confidence level is lower than a second threshold, the new lithology data set and the first large sample set are merged to obtain a second large sample set; Dividing the second largest sample set according to a preset second lithologic classification task to obtain a second task set; The improved metamodel is subjected to secondary inner and outer loop optimization according to the second task set to obtain an optimized metamodel.

6. A construction parameter optimization method based on a large model and meta-learning as claimed in claim 2, characterized in that: The optimized rock ballast data is segmented and labeled based on the primary SAM large model to obtain a labeled data set and prompt information, including: The primary SAM model is configured with an image encoder, a hint encoder, a mask generator, and a mask decoder to obtain a secondary SAM model; Identify and optimize the prompt information and multi-scale features in rock ballast data based on the secondary SAM model; The optimized rock ballast data are segmented and labeled according to the prompt information, multi-scale features and the secondary SAM model to obtain a labeled dataset.

7. A construction parameter optimization method based on a large model and meta-learning as claimed in claim 6, characterized in that: The rock slag volume features, rock slag boundary features and rock slag morphology features are calculated based on the preset three-dimensional reconstruction algorithm and segmentation mask set, including: Map all segmentation masks in the segmentation mask set to the three-dimensional point cloud space to obtain a valid point cloud subset; A three-dimensional shape model is created based on a three-dimensional reconstruction algorithm and a valid point cloud subset; The rock slag volume is calculated on the annotated dataset according to the rock slag volume calculation method and the three-dimensional morphology model to obtain the rock slag volume characteristics; The geometric characteristics of the rock slag volume characteristics are analyzed based on the three-dimensional morphological model to obtain the rock slag morphological characteristics and rock slag boundary characteristics.

8. The construction parameter optimization method based on large model and meta-learning according to claim 5, characterized in that: The performing of secondary inner and outer loop optimization on the improved meta-model according to the second task set to obtain an optimized meta-model includes: Obtaining a first training set from the second task set according to a preset second ratio; Performing inner loop optimization on the improved meta-model according to the first training set and a preset second gradient threshold to obtain a first task weight set and an adaptive meta-model; Obtaining a first validation set from the second task set according to a preset third ratio; Performing secondary inner and outer loop optimization on the adaptive meta-model based on the first task weight set and the first validation set to obtain a loss value and an enhanced meta-model; The global weights of the enhanced meta-model are updated according to the loss value to obtain the optimized meta-model.

9. The construction parameter optimization method based on large model and meta-learning according to claim 3, characterized in that: The performing of primary inner and outer loop training on the meta-learning model based on the first large sample set to obtain an empirical meta-model includes: Dividing the first large sample set according to the preset third lithologic classification task to obtain a third task set; Obtaining a second training set from the third task set according to a preset fourth ratio; Performing primary inner loop training on the meta-learning model according to the second training set and a preset third gradient threshold to obtain a primary meta-model; Obtain a second validation set from the third task set according to the preset fifth ratio; Evaluate the primary meta-model based on the second validation set to obtain evaluation results; The primary meta-model is trained in a primary outer loop according to the third task set and the evaluation results to obtain an empirical meta-model.

10. The construction parameter optimization method based on large model and meta-learning according to claim 2, characterized in that: The optimization processing of the preliminary rock slag data based on the preset multimodal data fusion model to obtain optimized rock slag data includes: The preliminary rock ballast data is fused through a multimodal data fusion model to obtain multimodal data; Obtain camera calibration data and laser scanner external parameters from real-time construction process; Perform unified mapping of multimodal data based on calibration data and external parameters to obtain aligned data; The multimodal data fusion model is used to optimize the aligned data to obtain optimized rock ballast data.

Images