A tunnel construction geological disaster dynamic identification method and system

Abstract

The present application belongs to the technical field of geological disaster identification, and provides a tunnel construction geological disaster dynamic identification method and system, geological and structural response data in the tunnel construction process are acquired, and the data is cleaned and feature enhanced; a dynamic horizontal extension network learning model is used to extract image primary fusion features of the data, the features are mapped to feature nodes through random weight mapping, and part of the feature nodes are randomly mapped to enhanced nodes; in response to changes in the tunnel construction conditions, an incremental learning mechanism is triggered, the feature nodes and the enhanced nodes are incremented, and the acquired data is newly added, the number of the feature nodes and the enhanced nodes is dynamically adjusted according to the feature distribution and complexity of the newly added data; the feature nodes and the enhanced nodes are taken as inputs, and a geological disaster dynamic identification model is used to perform real-time discrimination of disaster types and risk level evaluation. The present application can realize real-time perception and dynamic risk prevention and control of tunnel construction geological disasters.

Description

Technical Field

[0001] This invention belongs to the field of geological hazard identification technology, specifically relating to a method and system for dynamic identification of geological hazards during tunnel construction. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] The dynamic identification and prevention of geological hazards during construction has received widespread attention. When tunneling under complex geological conditions, dynamic processes such as the development of internal rock fissures, abrupt changes in seepage, and deterioration of surrounding rock play a fundamental and crucial role in assessing construction safety and providing early warnings of water inrush, mudslides, and collapse risks.

[0004] Modern tunnel construction has widely deployed intelligent sensing equipment such as video surveillance, 3D laser scanning, and multi-sensor systems, accumulating massive amounts of multi-source heterogeneous real-time monitoring data. However, current intelligent recognition methods based on deep learning, although highly accurate, generally suffer from bottlenecks such as high model training costs, difficulty in iterative updates, and difficulty in adapting to dynamically changing geological conditions during construction. The level of intelligence and the adaptability to engineering dynamics are seriously insufficient. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a method and system for dynamic identification of geological hazards during tunnel construction. This invention enables real-time perception and dynamic risk prevention and control of geological hazards during tunnel construction.

[0006] According to some embodiments, the present invention adopts the following technical solution: A method for dynamic identification of geological hazards during tunnel construction includes the following steps: Acquire geological and structural response data during tunnel construction, and perform data cleaning and feature enhancement. By using a dynamic laterally scaled network learning model, the primary image fusion features of the data are extracted, and then mapped to feature nodes through random weights. Some feature nodes are randomly mapped to enhancement nodes. In response to changes in tunnel construction conditions, an incremental learning mechanism is triggered to incrementally process feature nodes and enhancement nodes, and new data is acquired. The number of feature nodes and enhancement nodes is dynamically adjusted based on the feature distribution and complexity of the new data. Using feature nodes and enhancement nodes as input, a dynamic geological disaster identification model is used to perform real-time disaster type identification and risk level assessment.

[0007] As an alternative implementation method, the process of acquiring geological and structural response data during tunnel construction includes: acquiring video data of each key section of the tunnel, extracting key frame images from them, and preprocessing the images.

[0008] As an alternative implementation, the process of cleaning and enhancing the data includes: standardizing the original image, enhancing the overall and local contrast of the image by limiting contrast adaptive histogram equalization, removing image noise by using a combination of wavelet transform filtering and median filtering, and preserving edge details.

[0009] As an alternative implementation, the process of extracting primary image fusion features from data using a dynamic laterally scaled network learning model includes: the dynamic laterally scaled network learning model includes a primary feature extraction module and a dynamic laterally scaled learning network, wherein the primary feature extraction module uses a convolutional structure with fixed weights to extract primary image fusion features, each convolutional layer contains multiple convolutional kernels that can extract different image features, the obtained features are used as input to the dynamic laterally scaled learning network, and are mapped to feature nodes through random weights, and all feature nodes are further randomly mapped to enhancement nodes, wherein the number of enhancement nodes is determined by optimization using a Bayesian algorithm.

[0010] As a further implementation, the process of optimizing and determining the number of enhanced nodes using a Bayesian algorithm includes: using a Bayesian optimization method to search for parameters, constructing a probabilistic surrogate model, iteratively sampling in the hyperparameter space, updating the posterior distribution based on historical evaluation results, thereby guiding the search to converge toward a better region until the optimal parameter configuration is found.

[0011] As an alternative implementation, the process of triggering the incremental learning mechanism in response to changes in tunnel construction conditions includes: triggering the incremental learning mechanism when the recognition confidence level is lower than a set threshold for multiple consecutive recognitions, when a change in the construction stage or geological conditions is detected, and when a visual pattern that has not appeared in the training set is recognized and the number of consecutive occurrences of such a visual pattern exceeds a threshold.

[0012] As an alternative implementation, the process of incrementally adding feature nodes and enhancement nodes includes: assuming there are currently n feature nodes and m enhancement nodes, the (n+1)th feature node is generated by random mapping of the current image features, and a corresponding feature mapping enhancement node is added, or the (m+1)th independent enhancement node is added.

[0013] As an alternative implementation method, the process of dynamically adjusting the number of feature nodes and enhancement nodes based on the feature distribution and complexity of the newly acquired data includes: acquiring recent monitoring image data from the construction site, selecting qualified samples, and then standardizing, denoising, and enhancing the samples to form an incremental training sample set. Based on the incremental expansion mechanism of the dynamic horizontal expansion network learning model, the number of feature nodes and enhancement nodes is dynamically adjusted according to the feature distribution and complexity of the newly acquired samples. The output weights are updated using a recursive pseudo-inverse update algorithm, and the update formula is as follows: ; in, This is the feature mapping matrix corresponding to the newly added samples. Its corresponding label matrix, For the original weights, This is the intermediate update matrix obtained through recursive calculation.

[0014] As an alternative implementation method, the process of using a dynamic geological disaster identification model to identify disaster types and assess risk levels in real time includes: calculating a comprehensive risk value by comprehensively weighting four dimensions: identification confidence, disaster type, spatial location, and temporal evolution trend; and determining the corresponding risk warning level based on the risk threshold range in which the comprehensive risk value is located.

[0015] A dynamic identification system for geological hazards during tunnel construction includes: The data acquisition module is configured to acquire geological and structural response data during tunnel construction, and to clean and enhance the data. The dynamic lateral scaling module is configured to learn a model using a dynamic lateral scaling network, extract primary image fusion features from the data, map them to feature nodes through random weights, and randomly map some feature nodes to enhancement nodes. The incremental learning module is configured to respond to changes in tunnel construction conditions by triggering an incremental learning mechanism, incrementally processing feature nodes and enhancement nodes, and adding new data. Based on the feature distribution and complexity of the new data, the number of feature nodes and enhancement nodes is dynamically adjusted. The dynamic geological disaster identification module is configured to use feature nodes and enhancement nodes as inputs to perform real-time identification of disaster types and risk level assessment using the dynamic geological disaster identification model.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes an intelligent identification method for geological hazards in tunnel construction based on dynamic monitoring data. It inherits the advantages of dynamic horizontal expansion network learning models, such as fast training and support for incremental learning, and improves the discriminative representation of image features through high-performance lightweight convolutional modules. This provides an efficient and reliable technical path for real-time monitoring and early warning of tunnel construction hazards.

[0017] This invention establishes a dynamic monitoring dataset for geological hazards by real-time acquisition of geological and structural response data from multiple sensors during tunnel construction. To address the challenges of uneven spatiotemporal distribution of monitoring data, significant noise interference, and dynamic evolution of hazard patterns, data cleaning and feature enhancement strategies are introduced to improve the quality and representativeness of the input information. By constructing a dynamic, laterally expanded network learning model based on fused features and an incremental learning-based dynamic geological hazard identification model, the invention leverages the efficient feedforward computation capabilities and the incremental learning's ability to continuously adapt to new samples and evolution patterns, thereby enhancing the timeliness, accuracy, and dynamic updating capability of geological hazard identification in complex construction environments.

[0018] This invention enables real-time identification and dynamic risk assessment of disaster types such as collapses, water inrushes, rock bursts, and deformations. It integrates high-frequency monitoring data acquisition, rapid modeling through horizontally expanded network learning, and incremental model evolution. By comprehensively characterizing the tunnel state through multi-source features and dynamically adapting to geological changes through incremental learning, it forms a closed-loop system of perception-learning-early warning. It has the advantages of being lightweight, scalable, and capable of online learning, providing reliable technical support for real-time perception and dynamic risk prevention and control of geological disasters in tunnel construction. It has significant engineering application value and promising prospects for promotion.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0021] Figure 1 This is a schematic diagram illustrating the classification of tunnel construction hazards according to one embodiment. Figure 2 This is a schematic diagram of a dynamic laterally scaled network learning model based on fusion features, as exemplified by one embodiment. Figure 3 This is a schematic diagram of a method for dynamic identification of geological hazards during tunnel construction, according to one embodiment. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0026] Example 1 A method for dynamic identification of geological hazards during tunnel construction, such as Figure 3 As shown, it includes the following steps: Acquire geological and structural response data during tunnel construction, and perform data cleaning and feature enhancement. By using a dynamic laterally scaled network learning model, the primary image fusion features of the data are extracted, and then mapped to feature nodes through random weights. Some feature nodes are randomly mapped to enhancement nodes. In response to changes in tunnel construction conditions, an incremental learning mechanism is triggered to incrementally process feature nodes and enhancement nodes, and new data is acquired. The number of feature nodes and enhancement nodes is dynamically adjusted based on the feature distribution and complexity of the new data. Using feature nodes and enhancement nodes as input, a dynamic geological disaster identification model is used to perform real-time disaster type identification and risk level assessment.

[0027] The following is a detailed introduction.

[0028] First, it should be noted that the above method and process are practical application processes. Models such as dynamic horizontal scaling network learning models and dynamic geological disaster identification models all need to be constructed and pre-trained in advance. In order to make this solution clearer to those skilled in the art, the detailed process will be described in conjunction with the training process.

[0029] First, monitoring data is acquired. In this embodiment, high-definition visible light and infrared dual-mode cameras are fixedly deployed at multiple key sections of the tunnel, such as the tunnel face, support section, and seepage section, to achieve continuous image acquisition throughout the construction process.

[0030] Compared to traditional single monitoring methods or static data acquisition, high-definition video surveillance systems are more suitable for tracking the status of construction sites over large areas and continuously. They can clearly capture the visual characteristics and evolution of disasters such as landslides, water inrushes, and rock deformation. Different types of disasters in tunnel monitoring images exhibit distinct texture, shape, and motion characteristics, facilitating the monitoring of normal construction conditions and various disaster states by dynamic lateral expansion network learning models based on fused features and dynamic recognition models using incremental learning mechanisms.

[0031] The monitoring system, through continuous shooting and image extraction, provides real and continuous information on the temporal changes of disasters, enabling the construction of temporally correlated image sequence data. Explosion-proof high-definition cameras are deployed at key tunnel sections to capture continuous video streams during construction, from which keyframe images are extracted. Each image undergoes size standardization processing, such as... Figure 1 As shown, pixel-level annotation can be performed according to the three-level disaster classification system. During the annotation process, the main disaster area is distinguished from the background construction environment, forming high-quality image-label pairs. Finally, the complete annotated dataset is randomly divided into training, validation, and test sets in an 8:1:1 ratio, providing a standardized data foundation for subsequent model training and evaluation.

[0032] Next, we will perform data augmentation.

[0033] After constructing the tunnel construction disaster image dataset, systematic data augmentation is necessary to improve the model's generalization ability, robustness, and recognition accuracy. The original tunnel construction images are standardized, and contrast-limited adaptive histogram equalization enhances both overall and local contrast, making disaster features such as cracks, watermarks, and spalling areas more prominent. Since images are easily affected by dust, water mist, and equipment vibration during acquisition and transmission, a combination of wavelet transform filtering and median filtering is used to remove image noise while selectively preserving edge details to improve image quality. Secondly, different lighting conditions are simulated by adjusting image brightness and contrast, and geometric transformations such as random rotation, flipping, and cropping are employed to generate image samples with diverse perspectives. Constructing a more diverse image dataset effectively improves the model's generalization ability and prevents overfitting during training.

[0034] Construct and train a dynamic laterally scaled network learning model based on fusion features.

[0035] Compared to deep learning, the Dynamic Laterally Scaling Network (BLS) learning model offers an innovative learning strategy, primarily composed of feature nodes and augmentation nodes. Its network structure is simple, and its computational cost is significantly lower than that of deep learning. Input data is mapped to feature nodes using random weights, and these feature nodes are then mapped to augmentation nodes using random weights. Finally, the network output is estimated using the joint efforts of the feature nodes and augmentation nodes. However, directly applying BLS is insufficient for effectively extracting image features, necessitating the introduction of an image feature extraction module to compensate for BLS's shortcomings in image processing. Therefore, this invention incorporates a lightweight deep convolutional network with fixed weights before the dynamic laterally scaling network learning model to extract primary fusion features from tunnel disaster images, proposing a dynamic laterally scaling network learning model based on fusion features, as follows: Figure 2 As shown.

[0036] The model comprises two core components: a primary feature extraction module and a dynamically scaled-up learning network. The primary feature extraction module uses a fixed-weight convolutional structure to extract primary fusion features from the image, with each convolutional layer containing multiple kernels capable of extracting different image features. The resulting features serve as input to the dynamically scaled-up learning network and are mapped to feature nodes using random weights. To enhance the system's generalization ability, all feature nodes are further randomly mapped to augmentation nodes, with the number of augmentation nodes optimized using a Bayesian algorithm. Finally, these two types of nodes are used as common input to compute the output.

[0037] This embodiment utilizes a feature extraction module based on a lightweight deep convolutional network, employing the EfficientNet-B0 network structure pre-trained on the ImageNet dataset with fixed weights. This network achieves a good balance between computational efficiency and feature representation capability. The pre-trained weights on the ImageNet dataset ensure that the module has a general visual representation foundation, effectively avoiding the risk of overfitting in limited sample scenarios. Its lightweight characteristics are particularly suitable for the real-time dynamic recognition requirements in the complex environment of tunnel construction sites. The core component of the network, the MBConv module, integrates depthwise separable convolution and channel attention mechanisms, enabling it to adaptively focus on key local details and global contextual information in tunnel disaster images.

[0038] Suppose the input tunnel construction monitoring image is X The multi-level fusion features extracted by the EfficientNet-B0 module are as follows: ; in, This represents the extracted feature map. X The input is tunnel construction monitoring image data. This represents the set of all pre-trained and fixed parameters in the network. The forward propagation function of the entire network, composed of multiple MBConv modules, can be formally represented as: ; in, Indicates the first i Each MBConv module contains an operation sequence including expanded convolution, depthwise convolution, SE attention, and projective convolution. The resulting features... After global average pooling and flattening, a one-dimensional feature vector is formed. As input to a dynamic, laterally expanded learning network, feature nodes and enhancement nodes are generated through random mapping, ultimately enabling dynamic identification and classification of disasters such as collapses, water inrushes, and rock bursts during tunnel construction.

[0039] Unlike deep learning, dynamically scaled-out learning networks offer an innovative learning strategy, primarily composed of feature nodes and augmentation nodes. Their structure is simple, and their computational requirements are far lower than those of deep learning. First, input data is mapped to feature nodes using random weights. Second, feature nodes are mapped to augmentation nodes using random weights. Finally, feature nodes and augmentation nodes are used together to estimate the network's output.

[0040] Suppose the input data X contains N samples, each with an M-dimensional feature vector F. This feature vector is mapped to feature nodes using random weights: ; in, For the first i Each feature node and The weights and biases are randomized. n represents the number of feature nodes. Therefore, the j-th enhancement node is defined as: ; in, For the j-th enhancement node, and Let ξ be the random weights and biases, ξ be the non-linear activation function, and m be the number of augmentation nodes. = Let represent the combination matrix of all feature nodes; therefore, the model can be represented as: ; Where Y represents the model output matrix, [ | ] represents the combination matrix of feature nodes and augmentation nodes, and P is the output weight matrix. When the number of training samples is large, directly calculating the pseudo-inverse is costly. It can be approximated by ridge regression, as shown in the following formula: ; in, For regularization parameters, I Represents the identity matrix, [ | ] represents the combination matrix of feature nodes and enhancement nodes. Denotes the pseudo-inverse of the joint matrix. This represents the transpose of the joint matrix. It is the regularization matrix in ridge regression, which guarantees that the matrix inside the brackets is always invertible.

[0041] In a dynamically scaled learning network, the number of augmenting nodes is a hyperparameter, denoted as m, while the number of feature nodes is determined by the number of feature windows k and the number of nodes l in each feature window. To determine the optimal combination of hyperparameters, this embodiment employs a Bayesian optimization method for parameter search. This method constructs a probabilistic surrogate model to efficiently find the parameter configuration that optimizes model performance within the hyperparameter space.

[0042] In practical implementation, the Hyperopt package can be called, utilizing the tree-structured Parzen estimator as the core algorithm for Bayesian optimization. Hyperopt guides the search towards a better region by iteratively sampling in the hyperparameter space and updating the posterior distribution based on historical evaluation results. Compared to grid search or random search, Bayesian optimization can find superior hyperparameter combinations in fewer iterations. The core advantage of laterally scalable network learning is that it only requires computing the pseudo-inverse once, making it suitable for incremental learning and online updates. Its scalability can be improved by flexibly increasing the network width by adding k, l, and m, thereby enhancing accuracy.

[0043] In dynamic identification of tunnel construction disasters, to adapt to changes in geological conditions and continuous updates of image data during construction, the model's expressive power is expanded by dynamically adding feature nodes and enhancement nodes without retraining the entire network. The dynamic lateral expansion architecture based on the dynamic lateral expansion network learning model includes three incremental update modes: feature node increment, enhancement node increment, and input data increment, corresponding to three levels: model structure optimization, nonlinear expressive power enhancement, and new sample knowledge fusion, respectively.

[0044] First, let's introduce the incremental features and augmentation nodes: Let the initial dynamic horizontal scaling network learning model's combination matrix A consist of n feature nodes and m enhancement nodes extracted from EfficientNet-B0. When tunnel construction enters a new geological section, and the existing feature nodes are insufficient to represent new hazards such as new water inrush patterns or rockburst features, the node increment mechanism is automatically triggered. The newly added (n+1)th feature node is generated by randomly mapping the current image features F, as expressed by the formula: ; in, To add the (n+1)th feature node, , For randomly initialized weights and biases, This is a non-linear activation function. The corresponding newly added feature map enhancement node group is: ; in, For the reason The mapping generates a set of enhanced nodes, where ξ is a non-linear activation function. k=1,2…m, represents the random weights of the k-th augmentation mapping. , k=1,2…m, represents the random bias of the k-th augmentation mapping.

[0045] Of course, you can also directly add a (m+1)th independent enhancement node: in, Let ξ be the (m+1)th enhancement node, and let ξ be the nonlinear activation function. The feature node matrix, Let m+1 be the random weight matrix of the augmentation mapping. Let be the random bias vector of the (m+1)th augmentation mapping.

[0046] The updated combined matrix is ​​then expanded as follows: ; Where A′ is the expanded combination matrix, with new nodes appended to the right of the original A, and A is the original combination matrix. For newly added feature nodes, For newly added independent enhancement nodes, For the reason A set of enhanced nodes generated by the mapping.

[0047] By using the recursive update formula of the block matrix pseudo-inverse, the output weight P′ can be updated efficiently when only the pseudo-inverse corresponding to the newly added node is calculated, avoiding the need to re-invert the entire matrix. This significantly reduces the computational cost and time delay of model structure expansion and meets the real-time requirements of tunnel construction.

[0048] The process of inputting incremental data is as follows: Tunnel construction is a sequential process, and the amount of monitoring image data continuously increases over time, requiring support for incremental learning based on newly added samples. Let the set of newly added construction image samples be... Its corresponding feature vector is The newly added feature node mapping is as follows: ; in, This is the feature node matrix for the newly added samples. (This is for the newly added construction image sample set.) The n sets of feature nodes generated It is a non-linear activation function. To add a new set of construction image samples The corresponding feature vector, For the first The random weight matrix of the group of feature nodes, This is the bias term for the nth group of enhanced nodes.

[0049] The corresponding newly added enhanced node mappings are as follows: ; in, To target new samples The generated augmented node matrix contains m groups of augmented nodes, where ξ is a non-linear activation function. The feature node matrix of the newly added samples, Let be the random weight matrix of the i-th group of augmenting nodes. For the first Grouping enhances the bias term of nodes, increasing mapping flexibility. For the first The bias term of the group enhancement node, and The essence is the same, but the biases of different mapping layers are distinguished.

[0050] Let the new mapping matrix Then the expanded combination matrix is: ; Output weight updates can be efficiently accomplished through recursive pseudo-inverse calculation: ; in The labels are for the newly added samples. B is the intermediate matrix calculated according to the pseudo-inverse recursive update formula. This mechanism can quickly integrate new sample knowledge without retraining the entire network, maintaining model timeliness and recognition accuracy.

[0051] Incremental learning mechanism in dynamic identification of tunnel construction disasters: During tunnel construction, key construction-related indicators are monitored in real time. When an indicator meets a preset threshold, an incremental learning mechanism is automatically triggered, as follows: (1) The recognition confidence is consistently low. When the recognition confidence of multiple consecutive samples is lower than the dynamic threshold (which can be adaptively adjusted according to the construction stage), it indicates that the current model is not capable of feature extraction or classification of this type of image.

[0052] (2) Construction phase and geological condition changes. Combining construction logs and geological exploration data, when a new construction phase is detected (such as transitioning from hard rock excavation to soft surrounding rock section) or when significant changes occur in geological conditions, incremental learning is proactively initiated to adapt to the distribution of disaster characteristics in the new environment.

[0053] (3) Unknown visual patterns appear frequently. If a visual pattern that has not appeared in the training set is identified, and the number of consecutive occurrences of such a pattern exceeds a threshold, it is determined to be a new type of disaster or a new manifestation of a known disaster.

[0054] In the dynamic identification of tunnel construction disasters, once the incremental learning mechanism is triggered, the process will be executed to efficiently update and optimize the model. First, new data collection and preprocessing will be performed, automatically collecting recent monitoring image data from the construction site, screening out qualified samples, and then standardizing, denoising, and enhancing the samples to finally form the incremental training sample set. The process then proceeds to the dynamic node expansion and weight update stage. Based on the incremental expansion mechanism of the dynamically horizontally expanding network learning model, the number of feature nodes and augmentation nodes is dynamically adjusted according to the feature distribution and complexity of the newly added samples. On this basis, a recursive pseudo-inverse update algorithm is used to efficiently update the output weights, with the update formula as follows: ; in, This is the feature mapping matrix corresponding to the newly added samples. Its corresponding label matrix, For the original weights, This is the intermediate update matrix obtained through recursive computation. This mechanism only needs to calculate the pseudo-inverse of the newly added mapping part, avoiding retraining the entire model, thereby significantly improving learning efficiency and real-time response while ensuring recognition accuracy.

[0055] After constructing a dynamic identification and incremental learning mechanism for tunnel construction disasters, a dynamic monitoring and early warning mechanism for tunnel construction disasters based on a multi-level early warning strategy is established to achieve real-time perception and proactive prevention of disaster risks. First, the identification results of a dynamically and laterally expanded network learning model updated in real time are used as risk input. Then, after obtaining the comprehensive risk value and components of each dimension, a hierarchical monitoring and early warning mechanism is constructed.

[0056] (1) Multi-dimensional dynamic risk assessment Dynamic risk assessment calculations are performed based on a comprehensive determination of four dimensions: identification confidence level, disaster type, spatial location, and temporal evolution trend.

[0057] 1) Identify confidence dimensions This reflects the reliability of the model's current identification results. The lower the confidence level, the higher the uncertainty risk, which can be expressed by the formula: ; in, To identify the confidence risk value, Confidence is the model identification confidence, that is, the identification confidence of the output of the dynamically horizontally scaling network learning model, with a value range of [0,1].

[0058] 2) Disaster Type Dimension The disaster type dimension reflects the different potential hazards of different disaster types during tunnel construction. The type coefficient table, determined based on statistical analysis of tunnel construction accidents over the past decade and expert evaluation, is shown in Table 1. Table 1. Hazard Coefficient of Tunnel Disaster Types

[0059] 3) Spatial location dimension The spatial location dimension reflects the varying degrees of impact on construction safety and progress depending on the location of the disaster. Based on historical data statistics, engineering mechanism analysis, and expert experience assessment, the spatial location factor table is shown in Table 2. Table 2 Spatial Hazard Factors of Tunnel Construction Disasters

[0060] 4) Time Evolution Dimension The time evolution dimension reflects the speed trend of disaster development; the more obvious the trend, the higher the risk. Its quantitative expression formula is shown below: ; in, This represents the time trend risk value, a trend sensitivity coefficient, with a default value of 2.0, which can be adjusted according to the construction stage. The probability of disaster within a time window The change within, This refers to the observation time window.

[0061] The overall risk value is calculated by weighting the above four dimensions: ; in For the comprehensive risk value, i=1,2,3,4 are weight parameters and satisfy .

[0062] (2) Tiered early warning mechanism The comprehensive risk value calculated based on multi-dimensional dynamic risk assessment A scientific tiered early warning triggering mechanism has been established. This mechanism divides risk levels into four distinct levels: blue, yellow, orange, and red, each corresponding to different risk threshold ranges, visual color-coded indicators, and emergency response time requirements. Specifically, when 0.3 ≤ When the threshold is less than 0.5 (this threshold is based on the upper limit of the acceptable zone according to the risk acceptable criterion, obtained near the Youden index optimization point), a blue alert is triggered, indicating low risk, and on-site monitoring and patrols need to be strengthened within 30 minutes; when 0.5 ≤ A yellow alert is triggered when the risk level is less than 0.7 (this threshold is based on the lower quartile of the median risk of historical accident precursors; the risk boundary needs to be monitored). This indicates a moderate risk, requiring on-site verification and preliminary assessment to be completed within 15 minutes. When 0.7 ≤ When the threshold is less than 0.9 (this threshold is based on the upper limit of the acceptable risk criterion), an orange alert is issued, indicating a high-risk state, and a partial work stoppage and emergency response procedure must be initiated within 5 minutes; when When the threshold is ≥0.9 (based on the lower limit of the unacceptable zone), a red alert is issued, which means that the risk of disaster is extremely high and the highest level of emergency response, such as complete shutdown and evacuation of personnel, must be implemented immediately.

[0063] This grading mechanism, through quantified threshold definitions, intuitive color coding, and clear time constraints, achieves a structured connection from risk identification to early warning actions, providing clear and operable emergency guidelines for tunnel construction safety.

[0064] Example 2 A dynamic identification system for geological hazards during tunnel construction includes: The data acquisition module is configured to acquire geological and structural response data during tunnel construction, and to clean and enhance the data. The dynamic lateral scaling module is configured to learn a model using a dynamic lateral scaling network, extract primary image fusion features from the data, map them to feature nodes through random weights, and randomly map some feature nodes to enhancement nodes. The incremental learning module is configured to respond to changes in tunnel construction conditions by triggering an incremental learning mechanism, incrementally processing feature nodes and enhancement nodes, and adding new data. Based on the feature distribution and complexity of the new data, the number of feature nodes and enhancement nodes is dynamically adjusted. The dynamic geological disaster identification module is configured to use feature nodes and enhancement nodes as inputs to perform real-time identification of disaster types and risk level assessment using the dynamic geological disaster identification model.

[0065] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can be implemented in one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0066] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for dynamic identification of geological hazards during tunnel construction, characterized in that, Includes the following steps: Acquire geological and structural response data during tunnel construction, and perform data cleaning and feature enhancement. By using a dynamic laterally scaled network learning model, the primary image fusion features of the data are extracted, and then mapped to feature nodes through random weights. Some feature nodes are randomly mapped to enhancement nodes. In response to changes in tunnel construction conditions, when the identification confidence level falls below a set threshold for multiple consecutive identifications, detection... During the construction phase or when geological conditions change, and when visual patterns not present in the training set are identified, and such visual patterns... When the number of consecutive occurrences of a pattern exceeds a threshold, an incremental learning mechanism is triggered. Incremental processing is performed on feature nodes and enhancement nodes, and new data is added. The number of feature nodes and augmentation nodes is dynamically adjusted based on the feature distribution and complexity of the newly added data. Efficiently update output weights using a recursive pseudo-inverse update algorithm; by Dynamically adjusted Using feature nodes and enhancement nodes as input, a dynamic geological hazard identification model is used to perform real-time disaster type identification and risk level assessment. Based on four criteria: confidence level, disaster type, spatial location, and temporal evolution trend. A comprehensive dimensional assessment is used to perform dynamic risk evaluation and calculation. Based on the risk threshold range in which the comprehensive risk value falls, the corresponding risk level is determined. Risk warning classification .

2. The method for dynamic identification of geological hazards in tunnel construction as described in claim 1, characterized in that, The process of acquiring geological and structural response data during tunnel construction includes: acquiring video data of each key section of the tunnel, extracting key frame images from them, and preprocessing the images.

3. The method for dynamic identification of geological hazards in tunnel construction as described in claim 1, characterized in that, The process of cleaning and enhancing the features of the data includes: standardizing the original image, enhancing the overall and local contrast of the image by limiting the contrast adaptive histogram equalization, removing image noise by using a combination of wavelet transform filtering and median filtering, and preserving edge details.

4. The method for dynamic identification of geological hazards in tunnel construction as described in claim 1, characterized in that, The process of extracting primary image fusion features from data using a dynamically scaled horizontal network learning model includes: the dynamically scaled horizontal network learning model includes a primary feature extraction module and a dynamically scaled horizontal learning network. The primary feature extraction module uses a convolutional structure with fixed weights to extract primary image fusion features. Each convolutional layer contains multiple convolutional kernels that can extract different image features. The obtained features are used as input to the dynamically scaled horizontal learning network and are mapped to feature nodes through random weights. All feature nodes are further randomly mapped to enhancement nodes. The number of enhancement nodes is determined by optimization using a Bayesian algorithm.

5. The method for dynamic identification of geological hazards in tunnel construction as described in claim 4, characterized in that, The process of determining the number of augmenting nodes using Bayesian algorithm optimization includes: using Bayesian optimization method to search for parameters, constructing a probabilistic surrogate model, iteratively sampling in the hyperparameter space, updating the posterior distribution based on historical evaluation results, thereby guiding the search to converge toward a better region until the optimal parameter configuration is found.

6. The method for dynamic identification of geological hazards in tunnel construction as described in claim 1, characterized in that, The incremental process for feature nodes and enhancement nodes includes: assuming there are currently n feature nodes and m enhancement nodes, the (n+1)th feature node is generated by random mapping of the current image features, and a corresponding feature mapping enhancement node is added, or the (m+1)th independent enhancement node is added.

7. The method for dynamic identification of geological hazards in tunnel construction as described in claim 1, characterized in that, The process of dynamically adjusting the number of feature nodes and enhancement nodes based on the feature distribution and complexity of newly acquired data includes: acquiring recent construction site monitoring image data, screening qualified samples, and then standardizing, denoising, and enhancing the samples to form an incremental training sample set. Based on the incremental expansion mechanism of the dynamic horizontal scaling network learning model, the number of feature nodes and enhancement nodes is dynamically adjusted according to the feature distribution and complexity of the newly acquired samples. The output weights are updated using a recursive pseudo-inverse update algorithm, and the update formula is as follows: ； in, This is the feature mapping matrix corresponding to the newly added samples. Its corresponding label matrix, For the original weights, This is the intermediate update matrix obtained through recursive calculation.

8. A dynamic identification system for geological hazards during tunnel construction, characterized in that, include: The data acquisition module is configured to acquire geological and structural response data during tunnel construction, and to clean and enhance the data. The dynamic lateral scaling module is configured to learn a model using a dynamic lateral scaling network, extract primary image fusion features from the data, map them to feature nodes through random weights, and randomly map some feature nodes to enhancement nodes. The incremental learning module is configured as follows: In response to changes in tunnel construction conditions, when the identification confidence level is continuously recognized multiple times... When the degree is lower than the set threshold, when a construction phase or change in geological conditions is detected, or when a phenomenon not previously observed in the training set is identified. When a visual pattern is identified and its consecutive occurrences exceed a threshold, an incremental learning mechanism is triggered. Incremental processing is performed on feature nodes and enhancement nodes, and new data is added. Dynamically adjust features based on the feature distribution and complexity of the newly added data. The number of nodes and augmenting nodes is determined, and the output weights are updated efficiently using a recursive pseudo-inverse update algorithm. The dynamic geological disaster identification module is configured to... Dynamically adjusted Using feature nodes and enhancement nodes as input, a dynamic geological hazard identification model is used to perform real-time disaster type identification and risk level assessment. Based on identification confidence level, disaster Dynamic risk assessment is calculated by comprehensively determining four dimensions: type, spatial location, and temporal evolution trend. Determine the corresponding risk warning level based on the risk threshold range in which the value is located. .

Images