Underground engineering dynamic early warning method, electronic device and readable storage medium

Abstract

The application discloses a kind of underground engineering dynamic early warning method and electronic equipment, readable storage medium belong to underground space disaster prevention technical field.For the technical problem that underground environment interference is big and pure data driving deduces violation of physical law;First, the detection data of underground construction site is obtained, point cloud registration, optical flow tracking, space-time alignment and denoising cleaning are sequentially executed, and the spatial evolution characteristic parameter for representing rock mass deterioration trend is extracted;Second, after the parameter is fused with historical monitoring time sequence, it is input into the improved long-time sequence deduction network embedded with geomechanics prior matrix, and the future time sequence evolution trend curve of surrounding rock is calculated;Finally, relying on dynamic isolation forest model with time decay factor, dynamic safety boundary is autonomously learned and delimited, risk alarm level is determined, and is fed back to perception network to form closed loop.The application is mainly used for complex underground engineering disaster forward-looking early warning and construction safety dynamic intervention.

Description

Technical Field

[0001] This application relates to the field of disaster prevention and early warning technology in underground space, and in particular to a dynamic early warning method for underground engineering, electronic equipment, and readable storage medium. Background Technology

[0002] Under the combined effect of complex geological conditions and excavation unloading, the stress field and seepage field of underground rock and soil are prone to redistribution, which can induce engineering disasters such as large deformation of surrounding rock, rock bursts, and mud and water inrushes, posing a threat to the safety of on-site workers and construction equipment.

[0003] Currently, the safety inspection and control models commonly used in the engineering field rely heavily on regular manual observations and offline geological surveys. These methods not only suffer from significant time lags but also often produce isolated and low-frequency monitoring data. Faced with massive amounts of multimodal heterogeneous data, existing early warning systems frequently encounter the following technical bottlenecks: First, the underground environment is characterized by high dust levels and drastic changes in lighting, making traditional visual matching and point cloud registration algorithms prone to failure, generating a large amount of high-frequency data noise. Second, existing deep learning prediction models are mostly data-driven and lack constraints from geomechanical laws, making them prone to predictive divergence (overfitting or underfitting) that violate physical principles during long-term time-series simulations. Third, existing early warning systems often use fixed sensing frequencies and static thresholds, making it difficult to adapt to the dynamic evolution of the tunnel face as it continuously advances. Furthermore, they lack a closed-loop intervention mechanism of "prediction-feedback-sensing adjustment," leading to wasted system resources and a high false negative rate. Summary of the Invention

[0004] The purpose of this application is to provide a dynamic early warning method, electronic device, and readable storage medium for underground engineering, in order to solve the technical problems existing in related technologies, such as low noise resistance of data in harsh environments, poor physical consistency of long-term time-series inference, and difficulty in adapting to the dynamic evolution process of the tunnel face as it continues to advance when using fixed sensing frequencies and static thresholds.

[0005] According to a first aspect of the embodiments of this application, a dynamic early warning method for underground engineering is provided, comprising: The detection data collected by various sensors at the underground construction site includes three-dimensional point cloud data, advanced geological prediction radar map, surrounding rock deformation image and drilling data. The drilling data includes rock physical and mechanical parameters, three-dimensional characteristics of joint and fracture network and dynamic occurrence information of deep groundwater layer. Spatiotemporal alignment is performed on the three-dimensional point cloud data and the surrounding rock deformation image, and the advanced geological prediction radar image and the drilling data are denoised and cleaned to obtain the dynamic storage information of the deep groundwater layer after cleaning. Point cloud registration calculations were performed on the spatiotemporally aligned 3D point cloud data to obtain the relative convergence amount and settlement benchmark value. Optical flow tracing calculations were performed on the dynamic storage information of the cleaned deep groundwater layer and the spatiotemporally aligned surrounding rock deformation image to obtain the micro-slip vector; The relative convergence, settlement benchmark value and micro-slip vector are fused together, and combined with the physical and mechanical parameters of the rock mass and the three-dimensional characteristics of the joint and fracture network, spatial evolution characteristic parameters for characterizing the deterioration trend of the rock mass are extracted. Based on the aforementioned spatial evolution characteristic parameters and historical monitoring time series, an improved long-term time series extrapolation network with embedded geomechanical prior matrices is used to calculate the future time series evolution trend curve of the surrounding rock; Based on the future temporal evolution trend curve of the surrounding rock, a dynamic safety boundary is defined, and a risk alarm level is output.

[0006] Optionally, it also includes: Based on the risk alarm level, an early warning signal is issued, and at the same time, the status assessment result corresponding to the risk alarm level is fed back to the sensing network to dynamically and adaptively adjust the acquisition frequency and data transmission priority of various sensors, forming a closed-loop sensing and regulation.

[0007] Optionally, spatiotemporal alignment is performed on the three-dimensional point cloud data and the surrounding rock deformation image, and noise reduction and cleaning are performed on the advanced geological prediction radar image and the drilling data to obtain the dynamic storage information of the deep groundwater layer after cleaning, including: A three-dimensional geological space model is established based on the three-dimensional point cloud data and the physical and mechanical parameters of the rock mass. Combining multi-view stereo vision and close-range photogrammetry, spatiotemporal geometric consistency constraints are introduced to filter mismatched points, completing the three-dimensional reconstruction and spatial anchoring of the excavated surrounding rock. An instance segmentation network is used to perform pixel-level recognition and extraction of the three-dimensional features of the joint and fracture network mapped to the three-dimensional geological space model, resulting in spatiotemporally aligned three-dimensional point cloud data and surrounding rock deformation images. Spatial matching and cross-validation are performed between the advanced geological prediction radar map and the drilling data to identify and verify the hidden geological defects and water-rich structural zones ahead of the tunnel face. Based on the verification results, false abnormal signals and high-frequency noise in the detection data are filtered out to obtain the dynamic occurrence information of the deep groundwater layer after cleaning.

[0008] Optionally, point cloud registration calculations are performed on the spatiotemporally aligned 3D point cloud data to obtain the relative convergence and settlement benchmark values, including: By calculating the Gaussian curvature of the local surface of the spatiotemporally aligned 3D point cloud data, a first matching weight is assigned to structural feature regions where the curvature change exceeds a preset threshold, and a second matching weight is assigned to regions where the curvature change does not exceed the preset threshold. The first matching weight is greater than the second matching weight. The relative convergence of the cross section and the settlement benchmark value of the core region of the arch are calculated by iterating through a multi-scale voxel mesh from coarse to fine.

[0009] Optionally, optical flow tracing calculations are performed on the dynamic storage information of the cleaned deep groundwater layer and the spatiotemporally aligned surrounding rock deformation image to obtain minute slip vectors, including: An illumination-adaptive edge-preserving dense optical flow algorithm is used to process spatiotemporally aligned images of surrounding rock deformation and dynamic information of cleaned deep groundwater layers. Specifically, a local brightness conservation relaxation term and anisotropic diffusion regularization term are introduced into the optical flow energy functional to track the tiny slip vectors on the rock surface under illumination variation conditions.

[0010] Optionally, based on the spatial evolution characteristic parameters and historical monitoring time series, an improved long-term time series extrapolation network embedding a geomechanical prior matrix is ​​used to calculate the future temporal evolution trend curve of the surrounding rock, including: The displacement, deformation rate, and stabilization cycle sequence of the monitoring sections along the entire converging tunnel were collected as historical monitoring time series. Interpolation processing is performed on the non-equidistant missing data in the historical monitoring time series. Resampling and alignment compensation are completed by introducing the time dimension weight of the variogram function, thereby obtaining a multidimensional spatiotemporal sequence. The spatial evolution feature parameters are fused with the multidimensional spatiotemporal sequence, and the fused sequence is input into the improved Informer network constrained by physical information. Finally, a future temporal evolution trend curve of the surrounding rock with physical meaning is output. When performing temporal inference, the improved Informer network embeds the geomechanical prior matrix in its internal ProbSparse attention mechanism calculation.

[0011] Optionally, the computation process of the ProbSparse attention mechanism is as follows: After performing a dot product operation on the transpose of the query matrix and the key matrix, divide by the square root of the dimension scaling factor, and then subtract the penalty term obtained by multiplying the penalty coefficient, which is dynamically and adaptively adjusted according to the surrounding rock grade, with the geomechanical prior matrix. The subtraction result is then processed by a normalized exponential function, and finally the processed result is multiplied by the value matrix.

[0012] Optionally, a dynamic safety boundary is defined based on the future time-series evolution curve of the surrounding rock, and the corresponding risk alarm level is determined, specifically including: The future time-series evolution curve of the surrounding rock is input into a dynamic isolated forest model with a time decay factor to delineate dynamic safety boundaries and determine the corresponding risk warning levels; wherein: When constructing the detection tree, the dynamic isolated forest model assigns different attenuation weights to the normal surrounding rock deformation trajectory sample set according to the current spatial distance of the working face and the monitoring time span, so that the data that is closer to the current construction process and the newer the time has a greater influence on the boundary division. The dynamic isolated forest model autonomously learns the dynamic safety boundary under dynamically changing working conditions. When the predicted future temporal evolution trend curve of the surrounding rock exceeds the dynamic safety boundary, it determines the risk alarm level corresponding to the degree of danger.

[0013] According to a second aspect of the embodiments of this application, an electronic device is provided, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.

[0014] According to a third aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0015] The technical solutions provided by the embodiments of this application may include the following beneficial effects: This application achieves highly robust sensing in harsh underground environments. By performing spatiotemporal alignment of 3D point cloud data and surrounding rock deformation images, and by denoising and cleaning advanced geological prediction radar images and drilling data, and by performing point cloud registration and optical flow tracing calculations on this basis, the high-frequency noise caused by high-concentration dust in underground engineering and sudden changes in blasting illumination is effectively filtered out, ensuring the purity of the underlying data.

[0016] This application overcomes the bottleneck of evolution prediction in purely data-driven models. By employing an improved long-term inference network with embedded geomechanical prior matrices to calculate the future temporal evolution trend curve of the surrounding rock, the long-term prediction not only conforms to the statistical laws of data but also more rigorously adheres to the physical deformation boundaries of the rock mass, completely eliminating the divergence problem of deep network inference.

[0017] This application constructs a dynamic early warning standard that adapts to drift during tunneling. By defining dynamic safety boundaries based on the future temporal evolution curve of the surrounding rock and outputting risk alarm levels, it abandons the traditional static empirical thresholds, enabling the determination of safety boundaries to be dynamically updated to adapt to the progress of the tunnel face and the passage of time, greatly reducing the false alarm rate during complex geological transitions.

[0018] This application establishes a resource optimization closed loop of "perception-deduction-feedback". It innovatively uses risk assessment results to guide the perception network, automatically adjusting the sensor sampling frequency in high-risk areas. While ensuring high-frequency monitoring in the core area, it saves system transmission resources and realizes a closed loop of accurate judgment and automated control of geological disasters.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] Figure 1 This is a flowchart illustrating the overall process architecture of dynamic early warning for underground engineering projects with a closed-loop feedback mechanism, as presented in an embodiment of the present invention.

[0022] Figure 2 This is a flowchart illustrating the operation of the data processing module in establishing a digital basic database in an embodiment of the present invention.

[0023] Figure 3 This is a flowchart illustrating the operation of the spatiotemporal alignment and quality verification mechanism for multimodal heterogeneous detection data in this embodiment of the invention.

[0024] Figure 4 This is a flowchart illustrating the process of extracting spatial evolution characteristic parameters of rock mass based on an anti-noise algorithm in an embodiment of the present invention.

[0025] Figure 5 This is a flowchart illustrating the execution of long-sequence deduction based on the improved Informer architecture according to physical priors in an embodiment of the present invention.

[0026] Figure 6 This is a flowchart illustrating anomaly detection based on a dynamic time-decay isolated forest in an embodiment of the present invention. Detailed Implementation

[0027] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0028] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0029] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0030] Figure 1 This is a flowchart illustrating a dynamic early warning method for underground engineering according to an exemplary embodiment, such as... Figure 1 As shown, the method may include the following steps: Step S1: Obtain detection data collected by various sensors at the underground construction site. The detection data includes three-dimensional point cloud data, advanced geological prediction radar map, surrounding rock deformation image and drilling data. The drilling data includes rock physical and mechanical parameters, three-dimensional features of joint and fracture network and dynamic occurrence information of deep groundwater layer. Specifically, this step is mainly accomplished through the collaboration of a sensing network and a data processing module to build a digital foundational database. For example... Figure 2 The diagram shown is a flowchart illustrating the operation of the data processing module in establishing a digital foundation database in an embodiment of the present invention. After the sensing network is activated, multimodal heterogeneous detection data is continuously collected in the confined underground environment, and the various types of data are classified and processed to ultimately construct a digital foundation database. To overcome the limitations of a single data source, the multimodal heterogeneous detection data of the present invention includes: 3D point cloud data: acquired through a 3D laser scanner or array photogrammetry equipment. This data, in the form of a high-density set of 3D coordinate points, accurately represents the true spatial geometry and over- and under-excavation volume of the working face and the surrounding rock after excavation.

[0031] Advanced geological prediction radar map: obtained by ground penetrating radar or tunnel seismic wave advanced prediction system (TSP), it characterizes the spatial distribution of hidden structures inside the unexcavated rock mass in front of the tunnel face, such as faults and water-rich fracture zones.

[0032] Surrounding rock deformation images: acquired continuously using high-definition industrial camera equipment. These images visually record the changes in the two-dimensional characteristics of the surrounding rock surface during excavation and support, providing continuous basic image support for subsequent tracking and calculation of minute slip vectors on the rock mass surface.

[0033] Drilling data: acquired through advanced geological drilling or measurement-while-drilling equipment. This data not only records direct physical and mechanical parameters of the rock mass such as drilling speed, slewing torque, drill bit thrust, and RQD (rock quality index) extracted from the core, but also extracts the three-dimensional characteristics of the joint and fracture network of the rock mass through borehole photography and logging methods, and obtains dynamic information on the occurrence of deep groundwater layers through borehole hydrological observation.

[0034] Step S2: Perform spatiotemporal alignment on the three-dimensional point cloud data and the surrounding rock deformation image, and perform noise reduction and cleaning on the advanced geological prediction radar image and the drilling data to obtain the dynamic storage information of the deep groundwater layer after cleaning. Specifically, the spatiotemporal alignment unit of the data processing module extracts multimodal heterogeneous probe data from the basic database and simultaneously performs spatiotemporal alignment and noise reduction / cleaning quality verification. For example... Figure 3 The diagram shows the operation flowchart of the spatiotemporal alignment and quality verification mechanism for multimodal heterogeneous detection data. The system processes the extracted data in two branches: a "spatiotemporal alignment processing unit" and a "noise removal and cleaning processing unit," and finally merges and outputs high-precision underlying data.

[0035] In the spatiotemporal alignment process: a three-dimensional geological space model is established based on the three-dimensional point cloud data and the physical and mechanical parameters of the rock mass collected in the field; by combining multi-view stereo vision (MVS) and close-range photogrammetry, spatiotemporal geometric consistency constraints are introduced to filter mismatched points, and the point cloud data and field images are registered with high precision to complete the three-dimensional reconstruction and spatial anchoring of the excavated surrounding rock; an instance segmentation network is used to perform pixel-level recognition and extraction of the three-dimensional features of the joint and fracture network mapped to the three-dimensional geological space model, thereby obtaining the spatiotemporally aligned three-dimensional point cloud data and surrounding rock deformation image.

[0036] In the noise reduction and cleaning process, the key focus is on spatial matching and cross-validation between the advanced geological prediction radar map and the drilling data. First, the radar map is used to delineate the preliminary spatial range of concealed adverse geological conditions. Then, this range is spatially matched with the depth ranges in the drilling data where "drilling speed increases / decreases" or "torque mutations" occur. Through a fusion mechanism of "geophysical guidance and geochemical calibration," concealed geological defects and water-rich structural zones ahead of the tunnel face, such as within the 10-30 meter advance section, are accurately identified and verified. Based on the verification results, false anomalies and high-frequency noise in the detection data are filtered out, ultimately obtaining the dynamic occurrence information of the deep groundwater layer after cleaning.

[0037] The above design was adopted primarily because underground engineering faces harsh, confined spatial environments with high dust levels and drastic changes in lighting. Traditional visual matching and point cloud registration algorithms are prone to failure, resulting in a large amount of high-frequency noise and mismatched points in the detection data. Simultaneously, single geophysical or drilling methods often have limitations and are prone to false anomalies. This step effectively filters out mismatched points in spatial registration by introducing spatiotemporal geometric consistency constraints. Furthermore, in the multimodal fusion and quality verification stages, an innovative cross-validation mechanism of "geophysical guidance and geochemical calibration" is adopted, allowing the spatial areal data from the advanced forecast radar image to corroborate the point / line data of direct drilling mechanical parameters. This mechanism can accurately identify hidden geological defects, effectively filter out false anomalies and high-frequency noise from single data sources, and completely solve the technical problem of low noise resistance in harsh environments, ensuring the purity of the underlying data for subsequent extraction of accurate spatial evolution feature parameters.

[0038] like Figure 4 The diagram shows the workflow for extracting spatial evolution feature parameters of rock masses based on an anti-noise algorithm. This workflow mainly includes point cloud registration calculation, optical flow tracing calculation, and subsequent multimodal feature standardization and deep fusion, corresponding to steps S3 to S5 in this embodiment.

[0039] S3: Perform point cloud registration calculations on the spatiotemporally aligned 3D point cloud data to obtain the relative convergence amount and settlement benchmark value; Specifically, in step S3, the data processing module calls the multi-scale normal distribution transformation (NDT) registration algorithm based on curvature feature weighting to process high-frequency point cloud data.

[0040] In point cloud registration processing: First, the Gaussian curvature of the local surface of the spatiotemporally aligned 3D point cloud data is calculated. Structural feature regions where the curvature variation exceeds a preset threshold, such as crack intersections, are assigned a first matching weight (high matching weight). Regions where the curvature variation does not exceed the preset threshold are assigned a second matching weight, where the first matching weight is greater than the second matching weight. Subsequently, through multi-scale voxel mesh iteration from coarse to fine, spatial random noise interference from dust is overcome, and the relative convergence of the cross-section and the settlement benchmark value of the core area of ​​the vault are calculated.

[0041] When calculating the maximum relative convergence of the cross-section: This embodiment introduces a computational model with a nonlinear activation function, and extracts corresponding variables based on the registered features; ReLU(x)=max(0,x) is a nonlinear linear rectified function, and its mathematical formula is: in, The maximum feature convergence peak value at the kth monitoring frequency obtained from the characterization solution; NThis represents the total number of opposing survey lines laid out along this cross section; μ For survey line index identification; and Respectively characterize the first μ The group of survey lines is currently in the [number]th [section]. k Cycle and the previous cycle (the first cycle) k The physical distance measurement value (-1 times); To assign a spatial deterioration penalty weight to the survey line (this weight is adaptively assigned based on the rock fracture zone level extracted from the aforementioned three-dimensional geological spatial model).

[0042] In calculating the settlement baseline value: This embodiment constructs a spatial distance attenuation weighted settlement model. Discrete point cloud parameters are extracted for the core area of ​​the arch, and the mathematical formula is as follows: in: To calculate the obtained first k Mean value of periodic settlement characteristics; M This represents the total number of spatially discrete point cloud samples within the core force domain. For the first v Real-time absolute elevation of each discrete sampling point; This refers to a stable benchmark elevation that has been pre-calibrated and stored in the underlying database; The Gaussian decay confidence weights are assigned larger weights the closer the sampling point is to the geometric extremum of the vault.

[0043] The above design was adopted primarily because the underground environment is highly dusty, making traditional point cloud registration algorithms prone to failure and generating a large amount of high-frequency data noise. By introducing the curvature feature-weighted NDT registration algorithm, the matching weight of structural feature regions with drastic curvature changes is enhanced in a targeted manner, effectively filtering out high-frequency noise interference caused by high-concentration dust in underground engineering projects and ensuring the purity of the underlying data.

[0044] Furthermore, the fundamental reason for introducing a nonlinear rectification function into the solution model of relative convergence is that when the survey line span exhibits an outward divergence, i.e., when the deformation does not compress into the tunnel, the output is forcibly truncated to 0, thus completely avoiding the potential for missed local disasters caused by the cancellation of positive and negative errors at the algorithm's underlying level. Introducing a spatial distance attenuation weighted settlement model and Gaussian attenuation confidence weights into the settlement benchmark value solution can significantly amplify minute precursors of high-risk local collapses, greatly enhancing the ability to perceive early deterioration trends of the surrounding rock.

[0045] Step S4: Perform optical flow tracing calculation on the dynamic storage information of the cleaned deep groundwater layer and the spatiotemporally aligned surrounding rock deformation image to obtain the micro-slip vector; Specifically, the data processing module utilizes an illumination-adaptive edge-preserving dense optical flow algorithm to process the spatiotemporally aligned surrounding rock deformation image and the dynamic storage information of the cleaned deep groundwater layer. During processing, to address the shortcomings of traditional optical flow calculations, this step introduces a local brightness conservation relaxation term and anisotropic diffusion regularization term into the optical flow energy functional. Through this algorithm, the system can overcome the interference of ambient lighting conditions and accurately track and extract the minute slip vectors on the rock surface under varying illumination conditions.

[0046] The above design is adopted primarily because the underground engineering site environment is complex, with drastic changes in underground lighting. Traditional visual matching algorithms are prone to failure and generate significant high-frequency data noise, especially during blasting operations or sudden changes in construction lighting. By introducing an illumination-adaptive edge-preserving dense optical flow algorithm and innovatively adding a local brightness conservation relaxation term to the traditional optical flow energy functional, this invention effectively overcomes the optical flow tracking failure problem caused by sudden changes in blasting or construction lighting. This not only achieves highly robust perception in harsh underground environments, effectively filters out high-frequency noise interference caused by sudden changes in lighting and ensures the purity of the underlying data, but also ensures that the system can accurately capture extremely small slip precursor vectors on the rock surface.

[0047] Step S5: The relative convergence amount, settlement benchmark value and micro-slip vector are fused together, and the spatial evolution characteristic parameters used to characterize the deterioration trend of the rock mass are extracted by combining the physical and mechanical parameters of the rock mass and the three-dimensional characteristics of the joint and fracture network. Specifically, the feature extraction unit of the data processing module receives the multi-source heterogeneous parameters calculated from the underlying layer and performs a standardization and deep fusion process of multimodal features: Standardization and fusion of dynamic deformation indices: For the obtained relative convergence, settlement benchmark value of the core area of ​​the arch, and small slip vector of the rock surface obtained based on optical flow tracing, since these three have differences in dimensions and physical properties, such as convergence and settlement values ​​being displacements in a specific direction, while slip vectors contain multi-directional vector characteristics, the system first uses normalization algorithms, such as Z-score standardization, to eliminate the influence of dimensions and standardize and fuse slip vectors, convergence extrema, and weighted settlement values ​​into a unified data space.

[0048] Extraction and embedding of prior geomechanical features: From the established three-dimensional geological spatial model and basic database, rock mass physical and mechanical parameters reflecting the essential properties of the rock mass are extracted, such as drilling speed and torque obtained through drilling and RQD index of extracted rock cores, as well as three-dimensional features of joint and fracture networks extracted through instance segmentation.

[0049] Generation of spatial evolution characteristic parameters: The standardized dynamic monitoring indicators are combined with the aforementioned static geomechanical a priori features through tensor splicing and dimensionality reduction, transforming them into a unified high-dimensional feature vector or matrix. This comprehensive variable is the spatial evolution characteristic parameter. This parameter is no longer a single displacement value, but a comprehensive tensor that combines the "apparent deformation manifestation" and "intrinsic mechanical resistance boundary" of the surrounding rock, which can accurately characterize the current stress state and structural deterioration trend of the surrounding rock.

[0050] The above design is adopted primarily because, under the coupled effects of complex geological conditions and excavation unloading, the instability and failure of underground surrounding rock is not solely determined by displacement, but rather by the combined effects of dynamic deformation and the rock mass's own mechanical properties, such as hardness, integrity, and joints and fissures. Existing early warning systems often face the technical bottleneck of isolated monitoring data, tending to view deformation data in a singular and one-sided manner, lacking the constraints of geomechanical laws. This leads to the use of the same "absolute deformation scale" for judgment in different rock mass environments, such as brittle hard rocks prone to sudden rockbursts and weak surrounding rocks prone to slow large deformations, which easily results in false alarms or missed alarms.

[0051] Through this multimodal deep fusion design, the system breaks down data silos, tightly binding slip vectors reflecting microscopic faulting, convergence and settlement reflecting macroscopic displacement, with physical and mechanical parameters characterizing the rock mass's strength and fracture network. This design not only endows pure data with "physical meaning," successfully extracting spatial evolution characteristic parameters that truly reflect the essence of rock mass deterioration, but also lays a high-quality data foundation for embedding geomechanical prior matrices into the network for the next step of long-term time-series inference, completely eliminating the technical problem of pure data-driven models diverging from physical common sense in long-term time-series predictions.

[0052] Step S6: Based on the spatial evolution characteristic parameters and historical monitoring time series, an improved long-term time series extrapolation network embedding a geomechanical prior matrix is ​​used to calculate the future temporal evolution trend curve of the surrounding rock; this step includes the following sub-steps: S61: Collect the displacement, deformation rate, and stabilization cycle sequence of the entire monitoring section of the convergence tunnel as a historical monitoring time series. Specifically, the long-term time-series extrapolation unit of the intelligent decision-making module extracts prior historical data from the established basic database. For example... Figure 5 The diagram shows a flowchart of long-term time-series inference performed using an improved Informer architecture based on physical priors. Specifically, it involves collecting and converging historical displacement and deformation rates from each monitoring section along the entire tunnel, as well as the stabilization cycle sequence from deformation to reaching a stable state. This data set collectively constitutes the initial historical monitoring time-series sequence used by the long-term time-series inference network for learning.

[0053] S62: Interpolate the non-equidistant missing data in the historical monitoring time series, and complete resampling and alignment compensation by introducing the time dimension weight of the variogram function, thereby obtaining a multidimensional spatiotemporal sequence; Specifically, due to the complexity of the on-site construction environment, the acquired monitoring data often exhibits isolated and low-frequency characteristics, with non-equidistant missing data. The simulation unit first calls the adaptive spatiotemporal kriging interpolation algorithm to interpolate these non-equidistant missing data. During the interpolation process, by introducing the time dimension weight of the variogram, the missing data is resampled and aligned to compensate, thereby transforming the irregular time-series data into a standardized multidimensional spatiotemporal sequence.

[0054] S63: The spatial evolution feature parameters are fused with the multidimensional spatiotemporal sequence, and the fused sequence is input into the improved Informer network constrained by physical information. Finally, the future temporal evolution trend curve of the surrounding rock with physical meaning is output. When the improved Informer network performs temporal inference, the geomechanical prior matrix is ​​embedded in the calculation of the ProbSparse attention mechanism inside it. Specifically, the extracted spatial evolution feature parameters are fused with the aligned and compensated multidimensional spatiotemporal sequence. The fused sequence is then input into a physically constrained improved Informer network to perform long-term time-series inference. This improved Informer network embeds a geomechanical prior matrix based on rock mechanics equations into its core ProbSparse attention mechanism computation. M .

[0055] The computation process of the ProbSparse attention mechanism is as follows: After performing a dot product operation on the transpose of the query matrix and the key matrix, divide by the square root of the dimension scaling factor, and then subtract the penalty term obtained by multiplying the penalty coefficient (which is dynamically and adaptively adjusted according to the surrounding rock grade, which is evaluated by the aforementioned rock mass physical and mechanical parameters) with the geomechanical prior matrix. The above subtraction result is then processed by the normalized exponential function (Softmax), and finally the processed result is multiplied by the value matrix.

[0056] The above design is mainly adopted to solve the core technical bottleneck faced by existing early warning systems: most existing deep learning prediction models are purely data-driven and lack the constraints of geomechanical laws. In long-term time series extrapolation, they are prone to prediction divergence, overfitting or underfitting problems that violate physical common sense.

[0057] This step compensates for low-frequency isolated data using adaptive spatiotemporal kriging interpolation, ensuring the continuity of the sequence; the mathematical formula is: In the formula, Q , K , V For query, key-value matrix, d For dimensional scaling factor, λ This is a penalty coefficient that is dynamically and adaptively adjusted according to the surrounding rock grade.

[0058] More importantly, this application breaks through the bottleneck of evolutionary prediction in purely data-driven models and innovatively proposes an improved Informer network with physical information constraints. In the model calculation, when the deformation trend derived from pure underground engineering data exceeds the physical yield limit of the rock mass, the embedded geomechanical prior matrix... M This will forcefully suppress the abnormal weight. This design ensures that long-term time-series predictions not only conform to the statistical laws of massive data, but also more rigorously adhere to the physical deformation boundaries of the rock mass, completely eliminating the divergence problem in long-term time-series extrapolation of deep networks, and ultimately outputting a future time-series evolution trend curve of the surrounding rock with rigorous physical meaning.

[0059] Step S7: Determine the dynamic safety boundary based on the future time-series evolution trend curve of the surrounding rock and output the risk alarm level; Specifically, the anomaly detection unit of the intelligent decision-making module receives the physically meaningful evolution curve derived from the above deduction. For example... Figure 6 The diagram shows a flowchart for anomaly detection based on a dynamic time-decaying isolated forest. The future time-series evolution curve of the surrounding rock is input into a dynamic isolated forest model with a time-decay factor to define dynamic safety boundaries and determine the corresponding risk alarm level; where: When constructing the detection tree, the dynamic isolated forest model assigns different attenuation weights to the normal surrounding rock deformation trajectory sample set based on the current spatial distance to the working face and the monitoring time span. This mechanism ensures that data closer to the current construction process and more recent in time has a greater influence on boundary delineation, thereby enabling the safety boundary to dynamically and adaptively drift with the construction process.

[0060] The dynamic isolated forest model autonomously learns the dynamic safety boundary under dynamically changing working conditions. When the predicted future evolution trend curve of the surrounding rock exceeds this dynamic safety boundary, that is, when the evolution trend curve goes out of bounds, the system triggers a risk alarm signal and determines the risk alarm level corresponding to the degree of danger.

[0061] Specifically, the risk warning levels are classified from low to high based on the degree of deviation of the future time-series evolution curve of the surrounding rock from the dynamic safety boundary, the deformation rate, and the anomaly score output by the dynamic isolated forest model: 1. Blue Alert: A blue alert is issued when the evolution trend curve slightly exceeds the dynamic safety boundary (e.g., the deviation from the relative boundary is between 0% and 5%), and the anomaly score output by the dynamic isolated forest model is within a first preset range (e.g., the anomaly score is between 0.5 and 0.6). This indicates that the surrounding rock deformation is in an initial fine-tuning or slight slippage state, and the system issues a warning signal to remind on-site personnel to pay close attention to the cross-section.

[0062] 2. Yellow Alert: A yellow alert is issued when the evolution trend curve continuously exceeds the dynamic safety boundary and the deviation further increases (e.g., the deviation from the relative boundary is between 5% and 15%), and the anomaly score rises to the second preset range (e.g., the anomaly score is between 0.6 and 0.75). This indicates that the surrounding rock structure has undergone moderate deterioration, and joints and fractures may begin to expand at an accelerated pace. The system triggers conventional closed-loop sensing adjustment, automatically increasing the acquisition frequency of the sensors in this section.

[0063] 3. Orange Alert: An orange alert is issued when the evolution trend curve deviates significantly from the dynamic safety boundary (e.g., the deviation from the relative boundary is between 15% and 30%), the deformation rate remains high (e.g., the deformation rate exceeds 5 mm / d) and approaches the physical yield limit of the rock mass, and the anomaly score is in the third preset range (e.g., the anomaly score is between 0.75 and 0.9). This indicates that the surrounding rock has a significant risk of instability, large deformation, or collapse. The system immediately issues a serious warning to the on-site management terminal, prompting the need to take local support reinforcement measures.

[0064] 4. Red Alert: A red alert is issued when the evolution trend curve instantaneously or significantly crosses the dynamic safety boundary (e.g., the deviation from the relative boundary exceeds 30%), or when the predicted deformation directly exceeds the physical yield limit of the rock mass, and the anomaly score reaches or exceeds the extreme anomaly threshold (e.g., the anomaly score reaches the high-risk range of 0.9 to 1.0). This indicates a severe imbalance of internal stress in the surrounding rock structure, suggesting an imminent major engineering disaster such as large-scale collapse, mudslide, water inrush, or sudden rockburst. The system instantly issues the highest-level emergency rescue and evacuation signal.

[0065] The above design is adopted primarily because the working face of underground engineering is a constantly evolving, forward-moving process. Existing early warning systems mostly use fixed sensing frequencies and static thresholds for anomaly detection, which are difficult to adapt to such dynamically changing conditions and lack a feedback loop mechanism. When complex geological conditions change frequently, relying on static empirical thresholds can easily lead to a high false negative rate.

[0066] By introducing a dynamic isolated forest model with a time decay factor, this embodiment successfully constructs a dynamic early warning standard that adapts to drift during tunneling. This design completely abandons the traditional static empirical threshold, allowing the judgment weights of the safety boundary to be dynamically updated as the tunnel face advances and time progresses. This mechanism not only makes the early warning model more closely aligned with the latest actual working conditions but also significantly reduces the false alarm rate during complex geological transitions, providing a precise risk level determination basis for subsequent closed-loop feedback control.

[0067] To further achieve a closed-loop system for accurate assessment and automated management of geological hazards, and to maximize the conservation of system transmission and computing resources while ensuring accurate monitoring of high-risk core areas, the following may also be included: Step S8: Issue a warning signal based on the risk alarm level, and simultaneously feed back the status assessment result corresponding to the risk alarm level to the perception network to dynamically and adaptively adjust the acquisition frequency and data transmission priority of various sensors to form a closed-loop perception regulation.

[0068] Specifically, this invention introduces a dynamic adaptive acquisition mechanism: the sensing network does not operate at a fixed frequency, but receives in real time the "dynamic sensing compensation coefficient" issued by the backend system based on the closed-loop feedback of the state assessment results. For surrounding rock sections with different risk levels, the scanning and sampling frequency of point cloud and distributed optical fiber is adaptively increased or decreased to optimize system bandwidth resources.

[0069] The above design is adopted primarily because the monitoring data obtained by existing safety inspection and control models are often isolated and low-frequency. In addition, existing early warning systems mostly use fixed sensing frequencies and static thresholds, which are difficult to adapt to the dynamic evolution of the tunnel face as it continues to advance. Furthermore, they lack a closed-loop intervention mechanism of "prediction-feedback-sensing adjustment," which not only leads to a waste of system resources but also results in a high false negative rate.

[0070] By introducing the aforementioned multimodal and heterogeneous drilling data acquisition methods and dynamic adaptive sensing mechanisms, the system can use risk assessment results to guide the sensing network and automatically adjust the sensor sampling frequency in high- and medium-risk areas. This design ensures high-frequency monitoring of core dangerous rock areas while saving overall system transmission resources, thus successfully forming a closed loop for sensing resource optimization at the data source end, achieving accurate assessment and automated control of geological hazards.

[0071] As can be seen from the above embodiments, the dynamic early warning method for underground engineering proposed in this application effectively solves the core technical problems of traditional methods, such as low noise resistance of data in harsh environments, poor physical consistency of long-term time series extrapolation, and the difficulty of adapting to the dynamic evolution process of the tunnel face as it continues to advance, using fixed sensing frequencies and static empirical thresholds. Specifically, this embodiment achieves the following technical effects: Firstly, it achieves highly robust sensing in harsh underground environments. Faced with environments with high dust levels and drastic changes in illumination, this application utilizes multimodal heterogeneous detection data, namely cross-validation of geophysical and geochemical exploration data, and introduces curvature feature-weighted NDT registration and illumination-adaptive optical flow algorithms to effectively filter out high-frequency noise caused by environmental interference, ensuring the purity of the underlying data used for feature extraction.

[0072] Secondly, it breaks through the bottleneck of long-term evolution prediction in purely data-driven models. This application innovatively proposes an improved Informer network with physical information constraints, embedding geomechanical equations as prior matrices into the network's attention calculation mechanism. This makes the system's long-term predictions not only conform to the statistical laws of massive data, but also more rigorously adhere to the physical deformation boundaries of rock masses, completely eliminating the divergence of deep networks that violate physical common sense, i.e., the problems of overfitting or underfitting.

[0073] Third, a dynamic early warning standard that adapts to the tunneling progress was constructed. The traditional static empirical threshold was abandoned, and a dynamic isolated forest algorithm with a time decay factor was adopted. This allows the judgment weights of the safety boundary to be dynamically updated as the tunnel face advances and time progresses, greatly reducing the false alarm rate of early warnings when complex geological conditions frequently change.

[0074] Fourth, a closed-loop resource optimization mechanism of "perception-deduction-feedback" has been formed. Innovatively, the risk assessment and alarm results from the backend are fed back to the frontend, automatically adjusting the sensor sampling frequency in high- and medium-risk areas. While ensuring high-frequency and accurate monitoring of core dangerous rock areas, invalid high-frequency sampling in normal sections is avoided, significantly saving network transmission bandwidth and computing resources overall. Ultimately, this achieves dynamic, forward-looking early warning and closed-loop intervention for the safety of complex underground engineering construction.

[0075] Accordingly, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the above-described dynamic early warning method for underground engineering. The hardware structure of any data processing device in which the dynamic early warning method for underground engineering provided in this embodiment is implemented, in addition to including a processor, memory, DMA controller, disk, and non-volatile memory, may also include other hardware depending on the actual function of the data processing device, which will not be elaborated further.

[0076] Accordingly, this application also provides a computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, implement the aforementioned dynamic early warning method for underground engineering. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0077] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0078] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A dynamic early warning method for underground engineering projects, characterized in that, include: The detection data collected by various sensors at the underground construction site includes three-dimensional point cloud data, advanced geological prediction radar map, surrounding rock deformation image and drilling data. The drilling data includes rock physical and mechanical parameters, three-dimensional characteristics of joint and fracture network and dynamic occurrence information of deep groundwater layer. Spatiotemporal alignment is performed on the three-dimensional point cloud data and the surrounding rock deformation image, and the advanced geological prediction radar image and the drilling data are denoised and cleaned to obtain the dynamic storage information of the deep groundwater layer after cleaning. Point cloud registration calculations were performed on the spatiotemporally aligned 3D point cloud data to obtain the relative convergence amount and settlement benchmark value. Optical flow tracing calculations were performed on the dynamic storage information of the cleaned deep groundwater layer and the spatiotemporally aligned surrounding rock deformation image to obtain the micro-slip vector; The relative convergence, settlement benchmark value and micro-slip vector are fused together, and combined with the physical and mechanical parameters of the rock mass and the three-dimensional characteristics of the joint and fracture network, spatial evolution characteristic parameters for characterizing the deterioration trend of the rock mass are extracted. Based on the aforementioned spatial evolution characteristic parameters and historical monitoring time series, an improved long-term time series extrapolation network with embedded geomechanical prior matrices is used to calculate the future time series evolution trend curve of the surrounding rock; Based on the future temporal evolution trend curve of the surrounding rock, a dynamic safety boundary is defined, and a risk alarm level is output.

2. The method according to claim 1, characterized in that, Also includes: Based on the risk alarm level, an early warning signal is issued, and at the same time, the status assessment result corresponding to the risk alarm level is fed back to the sensing network to dynamically and adaptively adjust the acquisition frequency and data transmission priority of various sensors, forming a closed-loop sensing and regulation.

3. The method according to claim 1, characterized in that, The three-dimensional point cloud data and the surrounding rock deformation image are spatiotemporally aligned, and the advanced geological prediction radar image and the drilling data are denoised and cleaned to obtain the dynamic storage information of the deep groundwater layer after cleaning, including: A three-dimensional geological space model is established based on the three-dimensional point cloud data and the physical and mechanical parameters of the rock mass. Combining multi-view stereo vision and close-range photogrammetry, spatiotemporal geometric consistency constraints are introduced to filter mismatched points, completing the three-dimensional reconstruction and spatial anchoring of the excavated surrounding rock. An instance segmentation network is used to perform pixel-level recognition and extraction of the three-dimensional features of the joint and fracture network mapped to the three-dimensional geological space model, resulting in spatiotemporally aligned three-dimensional point cloud data and surrounding rock deformation images. Spatial matching and cross-validation are performed between the advanced geological prediction radar map and the drilling data to identify and verify the hidden geological defects and water-rich structural zones ahead of the tunnel face. Based on the verification results, false abnormal signals and high-frequency noise in the detection data are filtered out to obtain the dynamic occurrence information of the deep groundwater layer after cleaning.

4. The method according to claim 1, characterized in that, Point cloud registration calculations are performed on the spatiotemporally aligned 3D point cloud data to obtain the relative convergence and settlement benchmark values, including: By calculating the Gaussian curvature of the local surface of the spatiotemporally aligned 3D point cloud data, a first matching weight is assigned to structural feature regions where the curvature change exceeds a preset threshold, and a second matching weight is assigned to regions where the curvature change does not exceed the preset threshold. The first matching weight is greater than the second matching weight. The relative convergence of the cross section and the settlement benchmark value of the core region of the arch are calculated by iterating through a multi-scale voxel mesh from coarse to fine.

5. The method according to claim 1, characterized in that, Optical flow tracing calculations were performed on the dynamic storage information of the cleaned deep groundwater layer and the spatiotemporally aligned surrounding rock deformation image to obtain minute slip vectors, including: An illumination-adaptive edge-preserving dense optical flow algorithm is used to process spatiotemporally aligned images of surrounding rock deformation and dynamic information of cleaned deep groundwater layers. Specifically, a local brightness conservation relaxation term and anisotropic diffusion regularization term are introduced into the optical flow energy functional to track the tiny slip vectors on the rock surface under illumination variation conditions.

6. The method according to claim 1, characterized in that, Based on the aforementioned spatial evolution characteristic parameters and historical monitoring time series, an improved long-term time series extrapolation network embedding a geomechanical prior matrix is ​​used to calculate the future temporal evolution trend curve of the surrounding rock, including: The displacement, deformation rate, and stabilization cycle sequence of the monitoring sections along the entire converging tunnel were collected as historical monitoring time series. Interpolation processing is performed on the non-equidistant missing data in the historical monitoring time series. Resampling and alignment compensation are completed by introducing the time dimension weight of the variogram function, thereby obtaining a multidimensional spatiotemporal sequence. The spatial evolution feature parameters are fused with the multidimensional spatiotemporal sequence, and the fused sequence is input into the improved Informer network constrained by physical information. Finally, a future temporal evolution trend curve of the surrounding rock with physical meaning is output. When performing temporal inference, the improved Informer network embeds the geomechanical prior matrix in its internal ProbSparse attention mechanism calculation.

7. The method according to claim 6, characterized in that, The computation process of the ProbSparse attention mechanism is as follows: After performing a dot product operation on the transpose of the query matrix and the key matrix, divide by the square root of the dimension scaling factor, and then subtract the penalty term obtained by multiplying the penalty coefficient, which is dynamically and adaptively adjusted according to the surrounding rock grade, with the geomechanical prior matrix. The subtraction result is then processed by a normalized exponential function, and finally the processed result is multiplied by the value matrix.

8. The method according to claim 1, characterized in that, Based on the future time-series evolution curve of the surrounding rock, a dynamic safety boundary is delineated, and the corresponding risk warning level is determined, specifically including: The future time-series evolution curve of the surrounding rock is input into a dynamic isolated forest model with a time decay factor to delineate dynamic safety boundaries and determine the corresponding risk warning levels; wherein: When constructing the detection tree, the dynamic isolated forest model assigns different attenuation weights to the normal surrounding rock deformation trajectory sample set according to the current spatial distance of the working face and the monitoring time span, so that the data that is closer to the current construction process and the newer the time has a greater influence on the boundary division. The dynamic isolated forest model autonomously learns the dynamic safety boundary under dynamically changing working conditions. When the predicted future temporal evolution trend curve of the surrounding rock exceeds the dynamic safety boundary, it determines the risk alarm level corresponding to the degree of danger.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-8.

Images