Tunnel rock mass stability advance support method and system

Abstract

The application is a tunnel rock mass stability advanced support method and system, wherein a data synchronization fusion processing module calibrates time stamps and space coordinates of high-precision point cloud data and radar echo data, performs data fusion to generate a fused three-dimensional rock mass model, extracts a multi-dimensional fusion feature set, a working condition information acquisition and input module acquires current tunnel construction working condition information, an intelligent analysis and early warning module includes a space-time deep learning model unit and a dynamic threshold adjustment unit, the space-time deep learning model unit predicts rock mass deformation trend according to the multi-dimensional fusion feature set, the dynamic threshold adjustment unit dynamically corrects the early warning threshold according to the working condition information provided by the working condition information acquisition and input module, the intelligent analysis and early warning module generates a rock mass stability risk level based on the deformation trend prediction result and the dynamically corrected early warning threshold, and a support decision assistance module outputs a support scheme according to the rock mass stability risk level.

Description

Technical Field

[0001] This application relates to the field of tunnel engineering technology, specifically to a method and system for advanced support of tunnel rock mass stability. Background Technology

[0002] During tunnel construction, precise monitoring of the stability of the rock mass ahead and early implementation of advanced support are crucial to ensuring project safety. However, existing monitoring technologies suffer from systemic problems, making it difficult to meet the demands for high accuracy, immediacy, and reliability in safety predictions within complex geological formations. In terms of detection, most existing solutions rely on a single sensing technology, resulting in irreconcilable performance contradictions. While optical-based lidar can acquire high-precision microscopic deformations and fissure details on the rock surface, its detection performance deteriorates drastically in dusty, moisture-laden environments inside tunnels, and its effective detection range is very short. Millimeter-wave radar, based on microwave principles, offers excellent anti-interference capabilities and a longer detection range, revealing changes in the physical conditions deep within the rock mass. However, its spatial resolution is lacking, failing to accurately detect early surface damage signs such as microcrack formation. From the perspective of data analysis and early warning, existing methods generally rely on simple indicators (such as displacement) from a single data source to make static threshold judgments or use simple machine learning models. They lack the ability to perform deep fusion and intelligent analysis on multi-source, heterogeneous, and complementary data. As a result, their identification accuracy for complex hazard types such as gradual instability is low, with a high probability of errors and frequent omissions. In terms of system integration and engineering adaptation, the detection, processing, and early warning links are isolated from each other, lacking integrated hardware integration and data synchronization and fusion methods. This leads to slow system response, cumbersome operation, and fixed alarm logic, which cannot make real-time self-adjustments based on current geological parameters, construction methods, and other actual conditions, resulting in poor applicability in unpredictable underground environments. Summary of the Invention

[0003] To address the aforementioned issues, this application provides a method and system for advanced support of tunnel rock mass stability.

[0004] The technical solution of this invention is as follows:

[0005] A tunnel rock mass stability advanced support system includes:

[0006] The dual-radar fusion detection module includes a lidar unit and a millimeter-wave radar unit fixedly installed on the same mobile carrier platform. The lidar unit is used to collect high-precision point cloud data of the rock surface within its detection range, and the millimeter-wave radar unit is used to collect radar echo data that penetrates environmental interference media within its detection range.

[0007] The data synchronization and fusion processing module is communicatively connected to the dual radar fusion detection module. It is used to perform time stamp synchronization and spatial coordinate calibration on the high-precision point cloud data and radar echo data, and then perform data fusion to generate a fused three-dimensional rock mass model. It also extracts a normalized multi-dimensional fusion feature set from the fused three-dimensional rock mass model. The multi-dimensional fusion feature set includes: micro-deformation of the rock mass surface, fracture contour accuracy and surface texture change rate generated based on lidar data, and macro-displacement, deformation rate and echo energy change rate of the rock mass generated based on millimeter-wave radar data.

[0008] The working condition information acquisition and input module is used to acquire or receive the current working condition information of the tunnel construction in real time. The working condition information includes at least preset geological parameters characterizing the rock mass properties and identification information of the currently implemented construction method.

[0009] The intelligent analysis and early warning module, connected to the data synchronization and fusion processing module and the working condition information acquisition and input module, includes a spatiotemporal deep learning model unit and a dynamic threshold adjustment unit. The spatiotemporal deep learning model unit is used to predict the rock mass deformation trend within a preset time period based on the multi-dimensional fusion feature set of the time series. The dynamic threshold adjustment unit is used to dynamically correct the early warning threshold based on the working condition information provided by the working condition information acquisition and input module. The intelligent analysis and early warning module is used to generate a rock mass stability risk level based on the deformation trend prediction result and the dynamically corrected early warning threshold.

[0010] The support decision support module is used to match and output the corresponding support scheme from the preset support scheme library according to the rock mass stability risk level;

[0011] The design of the graded early warning system and visualization function module needs to be equipped with corresponding local sound and light warning devices and remote communication devices according to the rock mass stability level, while integrating three-dimensional geological model data and improving the presentation process of three-dimensional support structure information.

[0012] The integration of high-precision point cloud data and radar echo signal data within the preferred range should take into account tunnel dust content and relative humidity, and the fusion effect should be flexibly adjusted according to the importance coefficients of the two.

[0013] Preferably, the spatiotemporal deep learning model unit adopts a two-stream fusion neural network structure, including the following:

[0014] The time-series analysis branch uses the time-series data of the multi-dimensional fusion feature set as input to extract the temporal evolution characteristics of rock mass deformation;

[0015] The spatial analysis branch takes the data generated based on the fused three-dimensional rock mass model as input and is used to extract the spatial distribution characteristics of the rock mass structure.

[0016] The feature fusion and decision layer integrates the temporal evolution characteristics with the spatial distribution characteristics and then outputs the danger level of the rock mass's stability.

[0017] Preferably, the temporal analysis branch uses a long short-term memory network, and the input is a multi-dimensional fused feature set of N consecutive time periods, which is used to predict the deformation trend that may occur in the next M minutes; the spatial analysis branch uses a convolutional neural network, the purpose of which is to identify the fracture shape and structural texture from the data obtained from the above-mentioned fused three-dimensional rock mass model.

[0018] Preferably, the Long Short-Term Memory network is a two-layer stacked structure, where the first layer outputs complete temporal information and the second layer outputs the feature vector at the last moment, with a Dropout layer between them; the convolutional neural network has three or more convolutional layers using 3*3 convolutional kernels and their corresponding pooling layers.

[0019] Preferably, the output vector dimensions of the temporal analysis branch and the spatial analysis branch are the same, and channel splicing is performed at the feature fusion and decision layers.

[0020] Preferably, the feature fusion and decision layer includes a fully connected layer and a classification output layer, wherein the number of neurons in the classification output layer is equal to the number of types of rock mass stability risk levels.

[0021] Preferably, when the dynamic threshold adjustment unit is working, it pre-stores different benchmark threshold pairs corresponding to various working conditions, matches the corresponding benchmark threshold pairs with the real-time working conditions information obtained by the working conditions information acquisition and input module, and performs a dynamic correction operation between 0.7 and 1.3 based on the changes in the multi-dimensional fusion feature set at the current moment, thereby obtaining a new dynamically corrected warning threshold.

[0022] This application also provides a method for advanced support of tunnel rock mass stability, including:

[0023] S1. Control the lidar and millimeter-wave radar installed on the same mobile platform to simultaneously detect the excavated rock mass in the tunnel and acquire high-precision point cloud data and radar echo data respectively.

[0024] S2. After time-stamp synchronization and spatial coordinate calibration of the high-precision point cloud data and the radar echo data, they are fused to generate fused three-dimensional rock mass features.

[0025] S3. Extract normalized multi-dimensional fusion features from the fused three-dimensional rock mass model. The multi-dimensional fusion features include: micro-deformation of the rock mass surface, accuracy of fracture edge contour and surface texture change rate obtained based on lidar data, and macro-displacement, deformation rate and echo energy change rate of the rock mass obtained based on millimeter-wave radar data.

[0026] S4. Real-time acquisition or reception of current tunnel construction status information, wherein the status information includes at least preset geological parameters characterizing the rock mass properties and identification information of the currently implemented construction method.

[0027] S5. Based on the multi-dimensional fusion feature set of the time series, predict the rock mass deformation trend within a preset time period in the future;

[0028] S6. Dynamically correct the early warning threshold based on the working condition information provided by the working condition information acquisition and input module;

[0029] S7. Based on the deformation trend prediction results and the dynamically corrected early warning threshold, generate the rock mass stability risk level;

[0030] S8. Based on the rock mass stability risk level, match and output the corresponding support scheme from the preset support scheme library;

[0031] S9. Execute the corresponding local audible and visual alarm and remote information push according to the rock mass stability risk level, and visualize the fused three-dimensional rock mass model and the support scheme.

[0032] The beneficial effects of this invention are as follows:

[0033] After establishing a cyclical system encompassing dual-radar fusion detection, synchronous data processing, intelligent analysis and alerts, and dynamic support decision-making, the system achieved a core leap in tunnel rock mass stability monitoring: from passive response to proactive prediction, from a single judgment criterion to a combination of multiple sources, and from constant limits to dynamic adjustment. The system first uses lidar and millimeter-wave radar to complement each other, overcoming the shortcomings of individual sensors in terms of accuracy, interference resistance, and detection depth, thus forming a comprehensive 3D model including surface details and deep conditions. Based on this, the system employs a CNN-LSTM dual-stream deep learning model to simultaneously analyze the spatial structural hazards and temporal evolution trends of the rock mass, dynamically adjusting real-time conditions and optimizing early warning thresholds. This improves the accuracy of risk identification and the level of early warning. Finally, based on the determined risk level, the system automatically generates customized support plans with 3D spatial annotations and transmits them to frontline workers and remote management through layered alerts. The entire solution closely integrates detection, analysis, decision-making, and early warning, achieving intelligent monitoring and early warning, precise support decisions, and real-time risk response. It can effectively reduce the risk of collapse and safety expenses during tunnel construction, and significantly enhance the engineering safety assurance capability and construction efficiency under complex geological conditions. Attached Figure Description

[0034] Figure 1 This is a structural block diagram of a tunnel rock mass stability advanced support system according to an embodiment of this application;

[0035] Figure 2 This is a flowchart illustrating a method for advanced support of tunnel rock mass stability in an embodiment of this application. Detailed Implementation

[0036] Reference Figure 1 This application provides a tunnel rock mass stability advanced support system, including:

[0037] The dual-radar fusion detection module 101 includes a lidar unit and a millimeter-wave radar unit fixedly installed on the same mobile carrier platform. The lidar unit is used to collect high-precision point cloud data of the rock surface within its detection range, and the millimeter-wave radar unit is used to collect radar echo data that penetrates environmental interference media within its detection range.

[0038] The data synchronization and fusion processing module 102 is communicatively connected to the dual radar fusion detection module. It is used to perform time stamp synchronization and spatial coordinate calibration on the high-precision point cloud data and radar echo data, and then perform data fusion to generate a fused three-dimensional rock mass model. It also extracts a normalized multi-dimensional fusion feature set from the fused three-dimensional rock mass model. The multi-dimensional fusion feature set includes: rock mass surface micro-deformation, fracture contour accuracy and surface texture change rate generated based on lidar data, and rock mass macro-displacement, deformation rate and echo energy change rate generated based on millimeter-wave radar data.

[0039] The working condition information acquisition and input module 103 is used to acquire or receive the current working condition information of the tunnel construction in real time. The working condition information includes at least preset geological parameters characterizing the rock mass properties and identification information of the currently implemented construction method.

[0040] The intelligent analysis and early warning module 104 is connected to the data synchronization and fusion processing module and the working condition information acquisition and input module. It includes a spatiotemporal deep learning model unit and a dynamic threshold adjustment unit. The spatiotemporal deep learning model unit is used to predict the rock mass deformation trend within a preset time period based on the multi-dimensional fusion feature set of the time series. The dynamic threshold adjustment unit is used to dynamically correct the early warning threshold based on the working condition information provided by the working condition information acquisition and input module. The intelligent analysis and early warning module is used to generate a rock mass stability risk level based on the deformation trend prediction result and the dynamically corrected early warning threshold.

[0041] The support decision support module 105 is used to match and output the corresponding support scheme from the preset support scheme library according to the rock mass stability risk level;

[0042] The graded early warning and visualization module 106 enables local audible and visual alarms and remote information dissemination for the stability risk level of the rock mass, and displays the fused three-dimensional rock mass model and the support scheme.

[0043] In this embodiment, the mobile support platform adopts a tracked adaptive chassis, which is adapted to the rugged road surface in the tunnel and supports remote control or automatic tracking mode; the height of the platform is adjustable (the height can be adjusted, for example, between 1.5-3m), the dual radar integrated bracket can rotate 360° and tilt ±45°, and the installation distance between the lidar and millimeter-wave radar on the dual radar integrated bracket is fixed at 30cm, which meets the full coverage detection requirements of different construction sections.

[0044] During construction, operators activate the mobile support platform via remote control or a preset program, and the tracked adaptive chassis moves to the predetermined position on the working face. A telescopic structure driven by a servo motor is installed on the tracked adaptive chassis, extending and retracting longitudinally under the drive of the servo motor. A dual-radar integrated bracket is installed on top of the telescopic structure; specifically, this dual-axis rotatable gimbal allows for 360° rotation and ±45° pitch.

[0045] The specific structures of the tracked adaptive chassis, telescopic structure, and gimbal can all be implemented using existing technologies.

[0046] Inside the dual-radar fusion detection module, there is a high-precision synchronous clock transmitter (such as a GPS-tamed high-stability temperature-controlled crystal oscillator). This clock transmitter simultaneously sends a unified hardware trigger pulse (PPS signal) to the acquisition circuits of both the lidar unit and the millimeter-wave radar unit. Each time the hardware trigger pulse arrives, the two radars drive a data acquisition in strict synchronization, thereby ensuring that the two sets of data have the same time reference from the source and controlling the inherent time deviation between the two data streams to within 1 millisecond.

[0047] Hardware synchronization ensures that the acquisition of the two data streams begins at the same time. However, due to slight differences in subsequent processing and transmission to the data synchronization and fusion processing module, more accurate software alignment is required. After each round of hardware-triggered acquisition, the firmware of the LiDAR unit adds a microsecond-level timestamp T based on this high-precision synchronization clock to the acquired frame of point cloud data. Lidar The firmware of the millimeter-wave radar unit also adds the same timestamp T to each frame of radar echo data it acquires (usually a range-angle-complex amplitude matrix). MMLW Then, both radar units encapsulate their timestamps and data frames into a common standard format data packet and send it to the data synchronization and fusion processing module via network or bus.

[0048] The data synchronization and fusion processing module's buffer pool receives two radar data packet streams. For each arriving data packet, the module first analyzes its header to obtain the included timestamp. Then, it performs an instant matching algorithm: for each frame of LiDAR data (timestamp T... Lidar In terms of millimeter-wave radar data stream, searching for data that conforms to ||T MMW -T Lidar The matching frames with a time difference of | < δ (e.g., δ = 2 milliseconds) are selected, and the pair of data frames with the smallest time difference is used as tunnel images representing the same starting instant, thus forming a pair of data sets with spatiotemporal coordination characteristics.

[0049] After completing time synchronization and spatial coordinate correction, the data synchronization and fusion processing module will perform a deep integration operation on a pair of data in the same space-time (high-precision point cloud data from the lidar unit and radar echo data from the millimeter-wave radar unit). The key to this operation is to continuously adjust the proportion of high-precision point cloud data and radar echo data according to the changes in dust content and humidity inside the tunnel. The purpose is to overcome the problem of insufficient performance of a single sensor within a certain range and generate a highly accurate and reliable three-dimensional rock model after integration.

[0050] Before fusion, the data synchronization and fusion processing modules each extract complementary features that can be used for fusion from the two datasets. For the lidar point cloud, for each point or local patch, it calculates its three-dimensional coordinates, surface normal vector, local curvature value, reflection intensity, and information such as micro-deformation variables obtained from multiple data comparisons. These elements outline the detailed geometric structure and optical properties of the rock mass surface. As for each pixel in the millimeter-wave radar image (that is, representing a small area in a certain spatial orientation), it obtains the distance difference along the radar line of sight, deformation rate (that is, the change in deformation rate), backscattering attenuation, etc. These features show the changes and trends of the internal physical properties of the rock mass.

[0051] To achieve pixel-level / voxel-level precise fusion, the data synchronization fusion processing module creates a unified three-dimensional spatial index grid (voxel grid) that covers the entire detection area. This involves interpolating or assigning the attributes of each LiDAR feature point to a voxel, and then mapping the attributes of each millimeter-wave radar feature pixel to the corresponding voxel set according to its spatial coverage. This ensures that each three-dimensional voxel contains two feature vectors obtained from different radars.

[0052] The data synchronization and fusion processing module will calculate a set of two fusion weights for each individual element based on the current dust concentration and air humidity inside the tunnel, namely the lidar weight and the millimeter-wave radar weight.

[0053] When environmental conditions are favorable, the data synchronization and fusion processing module will assign a higher fusion weight to the lidar features in each voxel, such as WLidar=0.7, while assigning a lower fusion weight to the millimeter-wave radar features, such as WMMW=0.3.

[0054] When a significant increase in dust concentration or extremely high humidity is detected, the data synchronization and fusion processing module determines that the lidar signal has been severely interfered with. It then initiates automatic weight adaptation adjustment. At this point, the system dynamically reduces the fusion weights of each voxel based on its lidar features while increasing the corresponding millimeter-wave radar feature fusion weights. For each voxel, the adjustment range of its weights is directly proportional to the overall interference intensity detected in real time. Furthermore, normalization calculations are used to ensure that the sum of the two sets of weights in each voxel remains constant. This ensures that even in harsh environments, the fusion results of each voxel rely more heavily on millimeter-wave radar data with strong anti-interference capabilities. This effectively guarantees that the final fused 3D rock mass model has sufficient stability and reliability.

[0055] After weight allocation, the data synchronization fusion processing module performs fusion calculations for each voxel. The fusion calculation process is as follows: normalize the heterogeneous feature vectors of the two radars to the same range (0,1] to remove the influence of dimensions; calculate the final fused feature vector F of the voxel using the following formula. fused :

[0056]

[0057] in, and These are the normalized feature vectors of the lidar and millimeter-wave radar, respectively. By traversing all voxels and performing the above calculations, a new, unified, fused 3D rock mass model can be obtained. Each basic unit (voxel) in this fused 3D rock mass model contains multi-dimensional attributes of the radar data after fusion based on its reliability, such as high-precision surface geometry, reliable deep deformation, and changes in physical characteristics.

[0058] Find the corresponding points on the same rock mass surface on the existing fused 3D rock mass model and the historical fused 3D rock mass model, and calculate the distance between the two points as the distance difference in the surface normal direction. The difference is the micro deformation of the rock mass surface. Then divide the micro deformation of the rock mass surface by a preset maximum safety threshold and normalize it to 0-1. The larger the value, the more dangerous the deformation.

[0059] First, all points belonging to the same fracture are segmented from the integrated 3D rock mass model. Then, 3D plane fitting is performed on these points to calculate geometric parameters such as fracture opening, orientation, and length, which represent the fracture profile accuracy. These parameter values ​​are then divided by the reference values ​​for dangerous fractures set in engineering experience, and are normalized to a profile accuracy index between 0 and 1.

[0060] The surface texture change rate is indirectly characterized by analyzing the changes in laser reflection intensity on the rock surface in the fusion model. Specifically, the average reflection intensity of a fixed area at the current moment and the previous moment is calculated, and its change rate is determined. This change rate is then divided by the full range of laser intensity or a significant change threshold to obtain a normalized texture change rate index. This index can reflect surface conditions such as moisture, peeling, or contamination.

[0061] The macroscopic displacement of the rock mass is directly derived from the displacement value stored in each voxel of the fused 3D rock mass model, which is measured by millimeter-wave radar. Dividing this displacement value by a displacement warning threshold that is dynamically adjusted according to the current geological and construction conditions, the result is normalized to a macroscopic displacement index between 0 and 1.

[0062] In terms of time sequence, the macroscopic displacement of multiple consecutive frames in the same region is linearly fitted. The slope of the fitted line is the average deformation rate. This rate value is also divided by the dynamically adjusted rate warning threshold to complete the normalization.

[0063] By calculating the rate of change of millimeter-wave echo energy over time in a fixed region (echo energy variation rate), this change can reflect the alteration of water content or density within the rock mass. Dividing this rate of change by the background noise level under steady-state conditions yields the normalized echo energy variation rate index.

[0064] Ultimately, these six normalized independent metrics are combined into a single feature vector, i.e., a multi-dimensional fusion feature set.

[0065] In the working condition information acquisition and input module, the pre-set geological parameters are derived from the engineering exploration and design data before construction. This module, through a human-computer interaction interface, allows technicians to manually enter or import important geological parameters from files when starting the system or entering a new construction area. Parameters such as surrounding rock grade (e.g., II, III, IV), uniaxial compressive strength of the rock, rock integrity coefficient, and groundwater flow conditions are stored in the geological file corresponding to the current construction period and become fixed references for subsequent risk assessment. The foundation of this module is obtaining information related to the methods used in the current construction process, which requires technicians to manually select or input the corresponding methods when switching to a specific stage.

[0066] The spatiotemporal deep learning model unit in the intelligent analysis and early warning module is a dual-stream fusion neural network, including: a temporal analysis branch, which uses the aforementioned multi-dimensional fusion feature set temporal data as input to obtain the temporal evolution characteristics of rock mass deformation; a spatial analysis branch, which uses the data obtained by relying on the aforementioned fusion three-dimensional rock mass model as input to obtain the spatial distribution characteristics of rock mass structure; and a feature fusion and decision layer, which integrates the aforementioned temporal variation characteristics with spatial distribution characteristics and gives the risk level of the aforementioned rock mass stability.

[0067] Since the temporal analysis branch and the spatial analysis branch have the same terminal output dimension, the subsequent feature fusion stage and decision layer are processed using a channel-level splicing method.

[0068] The time series analysis branch mainly involves designing a Long Short-Term Memory (LSTM) network. It takes as input a comprehensive feature set of N dimensions spanning N consecutive defined periods to predict deformation over M minutes. The LTM network's input is a time data window of length N (e.g., n=80). The data at each moment consists of a high-dimensional fusion feature vector composed of six indicators (such as microscopic deformation and macroscopic displacement) that has been normalized before input.

[0069] Long Short-Term Memory (LSTM) networks have a two-layer structure. The first layer outputs the complete time series, and the second layer outputs the feature vector at the final time step. A Dropout layer is set between the two layers. Convolutional Neural Networks (CNNs) have at least three convolutional layers using 3x3 convolutional kernels and corresponding pooling layers.

[0070] The first LSTM layer of the Long Short-Term Memory (LSTM) network has 128 neurons and return_sequences is set to True. This layer outputs the complete hidden state sequence at each time step, thereby capturing short-term, local temporal correlations in the input sequence (such as small changes in deformation over a few minutes).

[0071] A Dropout layer with a dropout rate of 0.3 is added after the first layer. During training, a portion of neurons are randomly turned off, allowing the network to learn more robust features. This is an important means of preventing the model from overfitting to the training data.

[0072] The second LSTM layer has 64 neurons and `return_sequences=False`. This layer receives the sequence output from the first layer and outputs only the hidden state of the last time step. This hidden state vector contains the context of all long-term evolution of the N time step sequences and the final state, which can be regarded as a mathematical summary of the overall deformation trend of the rock mass over a period of time.

[0073] The hidden state vector output by the second LSTM layer is fed into a fully connected layer with 32 neurons and ReLU activation function to perform dimensionality reduction and feature enhancement processing, thereby generating the temporal feature vector V_temporal.

[0074] The spatial analysis branch employs a convolutional neural network (CNN) to identify fracture morphology and structural textures from the data obtained by fusing the 3D rock mass model. The three convolutional layers of the CNN use 32, 64, and 128 3x3 convolutional kernels, respectively. This combination of small 3x3 kernels effectively extracts spatial features from low to high dimensions: the first layer learns edges and corners (corresponding to the initial shape of micro-fractures); the second layer learns simple textures and shapes (corresponding to joints and bedding); and the third layer learns complex structural patterns (corresponding to potential slip surfaces and unstable rock mass outlines). Each layer is followed by a 2x2 max-pooling layer to reduce the spatial size of the feature map, allowing the neural network to focus more on spatial rather than positional changes. The result from the final convolutional module is transformed into a 128-dimensional feature vector through a global average pooling layer. This vector then passes through a fully connected layer with 32 neurons, forming the spatial feature vector Vspxral. Here, it is set to 32 dimensions, the same as VTEmorl, to ensure consistent weighting of information from both sides during the subsequent fusion process.

[0075] The feature fusion and decision layer is used to fuse the temporal evolution features and spatial distribution features, and output the rock mass stability risk level.

[0076] The feature fusion and decision layer comprises a fully connected layer and a classification output layer, which are interconnected. The number of neurons in the classification output layer is the same as the total number of rock mass stability risk level categories. After concatenating the two 32-dimensional vectors, V_temporal and V_spatial, through a channel concatenation operation, a 64-dimensional spatiotemporal fusion feature vector is formed. This channel-concatenated vector is fed into a two-layer fully correlated classification network. The first fully correlated layer has 64 neurons and uses the ReLU activation function to achieve a one-time superposition of spatiotemporal characteristics. The second fully correlated layer has 32 neurons and also uses ReLU activation. Finally, the number of neurons in the output corresponds to the number of each risk level (for example, four values ​​for levels I-IV).

[0077] In the entire intelligent analysis and prediction module, end-to-end training is conducted using labeled historical data, including dual-radar data and the final stability events. The loss function is generally in the form of cross-entropy loss. After training, the CNN stream can associate specific spatial patterns (such as clusters of cracks with special shapes) with high risk; while the LSTM stream learns to associate specific trend patterns (such as acceleration remaining consistently positive) with high risk. The decision layer also learns how to consider these two types of evidence. Even if the displacement does not change much (the risk level obtained from LSTM is relatively low), if the CNN detects a very poor crack structure (from which the risk level is high), the decision layer may still judge it as high risk. This is the important advantage of dual-stream integrated intelligence over single-indicator analysis.

[0078] The working process of the dynamic threshold adjustment unit includes: pre-storing benchmark threshold pairs corresponding to different working conditions; matching the corresponding benchmark threshold pairs according to the real-time working conditions information provided by the working conditions information acquisition and input module; and dynamically correcting the benchmark threshold pairs within a coefficient range of 0.7 to 1.3 based on the real-time fluctuation of the multi-dimensional fusion feature set, so as to generate the dynamically corrected warning threshold.

[0079] The benchmark threshold is a preset parameter corresponding to different working conditions. For example, a displacement warning threshold may have a benchmark pair of (Th_displacement_benchmark, Th_rate_benchmark). These come from three sources: first, the statistical results of data from many historical projects; second, the calculations or simulation experiments based on geotechnical mechanics theory; and third, the accumulated experience of experts in the field. This knowledge base uses lookup tables or parametric models to synthesize various lithologies, burial depths, and construction methods into the values ​​corresponding to specific benchmark thresholds, providing a preliminary reference standard for subsequent risk assessment.

[0080] During tunnel construction, the working condition information collection and input module provides the current working condition background. The dynamic threshold adjustment module, following this situation (such as Class III rock structure, deep burial conditions, or excavation using the CD method), quickly searches the database to find the optimal pair of benchmark parameter values ​​Th_base that best match the current construction stage. This process ensures that the initial settings of the alarm system are consistent with the actual geological characteristics and human operation procedures, thus resolving the problem that static values ​​alone cannot accurately reflect the current state of the project.

[0081] When data fluctuations are stable and the signal-to-noise ratio is high (e.g., a small standard deviation), it indicates a good monitoring environment and reliable data. In this case, the unit adopts a contraction strategy, multiplying the baseline threshold by a correction factor less than 1 to obtain a more sensitive alarm value: Th_alert = 0.8 × Th_base. This allows the system to detect smaller anomalies. When data fluctuates drastically and noise is significant (e.g., a large increase in standard deviation), it may be due to transient interference such as vibration from construction machinery or explosive impacts. To prevent false alarms, the unit adopts a tolerance strategy, multiplying the baseline threshold by a larger adjustment factor (e.g., 1.2) to obtain a broader warning limit: Th_alert = 1.2 * Th_base. This enhances the system's ability to cope with interference in complex environments. The positive coefficient is limited to between 0.7 and 1.3 to ensure the adjustment remains within a controllable range and does not deviate too far from engineering safety standards.

[0082] The finally generated dynamically corrected warning threshold Th_alert will serve as a reference for comparison and decision-making in the intelligent analysis and warning module, and will be used to compare with the deformation trend value predicted by the spatiotemporal deep learning model unit to determine its risk level.

[0083] The process of generating a rock mass stability risk level based on the deformation trend prediction results and the dynamically corrected warning thresholds includes: comparing one by one the multiple future deformation indicators (e.g., predicted displacement D_pred and predicted deformation acceleration A_pred) predicted by the spatiotemporal deep learning model unit with the corrected warning thresholds (e.g., displacement threshold Th_d and acceleration threshold Th_a) provided by the dynamic threshold adjustment unit. For each indicator i, its risk exceedance ratio R_i is calculated: R_i = Value_pred_i / Th_alert_i; if R_i < 1, it indicates that the predicted value of the indicator is within the safe range; if R_i ≥ 1, it indicates that the indicator is predicted to exceed the limit, and the exceedance state of each indicator constitutes a primary risk signal.

[0084] The spatial feature vectors given by the CNN or the semantic segmentation results of the hidden danger region (if there is a through crack) are transformed into a spatial risk coefficient S (for example, the value of S can be set between 0.8 and 1.5). Once the CNN detects a high-risk spatial structure, it will make S greater than 1, producing a risk amplifier effect. If there is no problem with the spatial structure, it will be approximately equal to 1 or less than 1.

[0085] The finally determined risk level is achieved by a comprehensive decision function that fuses the above information. A simplified decision model is as follows: The comprehensive risk value Risk_score = f(maxRi, S, trend continuity); maxRi represents the worst degree of exceeding the limit in the indicator prediction; S represents the inherent vulnerability of the current rock; trend continuity represents how long the predicted trend will continue before exceeding the limit again. (That is, the situation of exceeding the limit occurs continuously for several days).

[0086] The achievement of this function f can be either a pre-set weighted summation and logical rules, or a relatively lightweight machine learning classifier (such as a random forest). A risk level classification table is formulated in advance, and the calculated Risk_score is mapped to a specific risk level (such as levels I - IV).

[0087] Level I (Safe): Ri << 1, S normal, trend stable

[0088] Level II (Attention): It means that some of the Ri are approaching 1, such as 0.8 < Ri < 1 or S has risen slightly. At this time, attention should be paid in a timely manner.

[0089] Level III (Warning): At least one core indicator Ri ≥ 1 or S has risen significantly. Immediately issue an alarm and give prompts for corresponding preparatory work.

[0090] Level IV (Dangerous): When several core indicators with Ri > 1, S is very high, and there is an obvious accelerating deterioration trend, it reaches the highest alarm level. Construction must be stopped immediately, and emergency support measures must be taken.

[0091] In the hierarchical alarm and visualization section, first, according to the determined risk levels, the pre-set local acoustic and optical alarm schemes are activated. Such alarm devices are generally integrated into mobile carrying platforms or key locations in the tunnel. They use different colors, frequencies, and sound forms to represent the danger level. For example, a long blue light represents normal system operation, a flashing yellow light reminds to observe carefully, a rotating red light plus an intermittent alarm sound is a warning signal, and a flashing red light plus a continuous alarm sound is the most serious danger. In this way, in the noisy tunnel environment, the staff can immediately and clearly detect changes in the dangerous situation, and use the integrated radio communication module to send hierarchical information to remote receiving devices. The early warning information is定向 transmitted to each level of guardians in accordance with the degree of danger through the construction protection platform, exclusive mobile applications, or text messages. A complete early warning information covers both the danger level and the occurrence time point, and also includes the main judgment results given by the intelligent analysis section, such as the location of key hidden dangers and their estimated deformation trends. For relatively serious dangerous situations, the system will trigger continuous message transmission and confirm with the recipient, so as to ensure that important information is accurately delivered to the responsible person.

[0092] When information is pushed, this module dynamically updates in its visual interactive interface and highlights key areas. The newly formed fused 3D rock mass model is immediately displayed in the main areas. Then, based on the assessment of the degree of danger, it is drawn using bright colors, contour lines, or special graphics to clearly show the discovered risks and their corresponding danger levels. For example, a potential sliding crack that is classified as a level III risk will be highlighted in a conspicuous orange area.

[0093] Along with the display of risk areas, detailed information on recommended support schemes created by the support decision support module will also pop up on the interface. This information is presented using a combination of text and graphics, covering support types and design parameter lists. Crucially, the scope of the support scheme, such as anchor placement locations or steel frame assembly areas, is accurately attached to the 3D rock mass model using a semi-transparent overlay or vector image. This allows construction personnel to clearly understand where the dangers are and how to take appropriate countermeasures.

[0094] In addition, there are functions for historical data review and multi-period comparison. Users can view historical deformation curves of any potential hazard area, changes in risk level, previous support methods, etc., to comprehensively analyze the risk evolution process and support effectiveness. All visual information can be adjusted in terms of perspective, zoom in / out, and pan. During this process, users can observe the hazard situation and its response strategies from various directions, and can easily export the results in the form of reports or photos for use in meetings or on-site presentations.

[0095] By forming a closed-loop system that integrates dual-radar fusion detection, simultaneous data processing, intelligent analysis and alarm functions, and dynamic support decision-making, the detection method for tunnel rock mass stability has shifted from passive response to active prediction. Furthermore, it no longer relies on a single evaluation criterion but adopts a multi-factor collaborative judgment approach. Finally, it moves beyond pre-determined standard dimensions to a more flexible and adjustable standard range. The system first relies on the complementary use of lidar and millimeter-wave radar to overcome the limitations of individual sensors in terms of accuracy, anti-interference, and detection depth, thereby creating a comprehensive 3D fusion model that includes surface details and deep information. Based on this, the system utilizes a CNN-LSTM dual-channel deep learning architecture to comprehensively analyze the spatial distribution characteristics of the rock mass and its changes over time. Simultaneously, it uses operating parameters to adjust and improve pre-set alarm threshold values ​​in real time. This approach significantly improves the accuracy of potential disaster assessment and the speed of early warning response. The platform develops specialized support design proposals based on accurately assessed risk levels, displaying them in a three-dimensional coordinate system. A multi-layered early warning system ensures efficient message transmission, guaranteeing that critical data is immediately transmitted to frontline workers and senior management for processing. This integrated system combines detection, data processing, decision support, and early warning functions, enabling intelligent design of the monitoring system, refined adjustment of support strategies, and optimized timeliness of emergency response. This effectively reduces the probability of collapse and safety costs in tunnel engineering, significantly improving construction efficiency and overall safety under complex geological conditions.

[0096] according to Figure 2 As shown, the embodiments of this application propose a method for improving tunnel rock mass stability and carrying out preliminary support:

[0097] S1. Control the lidar and millimeter-wave radar installed on the same mobile platform to simultaneously detect the excavated rock mass in the tunnel and acquire high-precision point cloud data and radar echo data respectively.

[0098] S2. After time-stamp synchronization and spatial coordinate calibration of the high-precision point cloud data and the radar echo data, they are fused to generate fused three-dimensional rock mass features.

[0099] S3. Extract normalized multi-dimensional fusion features from the fused three-dimensional rock mass model. The multi-dimensional fusion features include: micro-deformation of the rock mass surface, accuracy of fracture edge contour and surface texture change rate obtained based on lidar data, and macro-displacement, deformation rate and echo energy change rate of the rock mass obtained based on millimeter-wave radar data.

[0100] S4. Real-time acquisition or reception of current tunnel construction status information, wherein the status information includes at least preset geological parameters characterizing the rock mass properties and identification information of the currently implemented construction method.

[0101] S5. Based on the multi-dimensional fusion feature set of the time series, predict the rock mass deformation trend within a preset time period in the future;

[0102] S6. Dynamically correct the early warning threshold based on the working condition information provided by the working condition information acquisition and input module;

[0103] S7. Based on the deformation trend prediction results and the dynamically corrected early warning threshold, generate the rock mass stability risk level;

[0104] S8. Based on the rock mass stability risk level, match and output the corresponding support scheme from the preset support scheme library;

[0105] S9. Execute the corresponding local audible and visual alarm and remote information push according to the rock mass stability risk level, and visualize the fused three-dimensional rock mass model and the support scheme.

[0106] It should be understood that the application of this application is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Those skilled in the art can understand that implementing all or part of the processes of the above embodiments and making equivalent changes according to the claims of this application still fall within the scope of this application.

Claims

1. A tunnel rock mass stability advanced support system, characterized in that, include: The dual-radar fusion detection module includes a lidar unit and a millimeter-wave radar unit fixedly installed on the same mobile carrier platform. The lidar unit is used to collect high-precision point cloud data of the rock surface within its detection range, and the millimeter-wave radar unit is used to collect radar echo data that penetrates environmental interference media within its detection range. The data synchronization and fusion processing module is communicatively connected to the dual radar fusion detection module. It is used to perform time stamp synchronization and spatial coordinate calibration on the high-precision point cloud data and radar echo data, and then perform data fusion to generate a fused three-dimensional rock mass model. It also extracts a normalized multi-dimensional fusion feature set from the fused three-dimensional rock mass model. The multi-dimensional fusion feature set includes: micro-deformation of the rock mass surface, fracture contour accuracy and surface texture change rate generated based on lidar data, and macro-displacement, deformation rate and echo energy change rate of the rock mass generated based on millimeter-wave radar data. The working condition information acquisition and input module is used to acquire or receive the current working condition information of the tunnel construction in real time. The working condition information includes at least preset geological parameters characterizing the rock mass properties and identification information of the currently implemented construction method. The intelligent analysis and early warning module, connected to the data synchronization and fusion processing module and the working condition information acquisition and input module, includes a spatiotemporal deep learning model unit and a dynamic threshold adjustment unit. The spatiotemporal deep learning model unit is used to predict the rock mass deformation trend within a preset time period based on the multi-dimensional fusion feature set of the time series. The dynamic threshold adjustment unit is used to dynamically correct the early warning threshold based on the working condition information provided by the working condition information acquisition and input module. The intelligent analysis and early warning module is used to generate a rock mass stability risk level based on the deformation trend prediction result and the dynamically corrected early warning threshold. The support decision support module is used to match and output the corresponding support scheme from the preset support scheme library according to the rock mass stability risk level; The graded early warning and visualization module is used to execute corresponding local audible and visual alarms and remote information push according to the rock mass stability risk level, and to visualize the fused three-dimensional rock mass model and the support scheme.

2. The system according to claim 1 is characterized in that the data fusion process of high-precision point cloud data and radar echo data in the data synchronization fusion processing module is to appropriately adjust the weight values ​​of high-precision point cloud data and radar echo data according to the amount of dust and humidity in the tunnel.

3. Based on the technical solution described in claim 1, the mentioned spatial temporal deep learning model unit constructs a neural network structure framework based on the dual-stream fusion theory; The time series analysis module explores the evolutionary characteristics and dynamic laws of rock strata deformation by constructing time series data based on multidimensional integrated feature sets. The spatial analysis branch uses the data obtained from the fusion of the three-dimensional rock mass model as input to obtain the spatial distribution characteristics of the rock mass structure. The feature fusion and decision layer is used to integrate the mentioned temporal evolution characteristics with spatial distribution characteristics before giving the corresponding rock mass stability hazard level.

4. The system described in claim 3 is characterized in that the system's temporal analysis branch uses a long short-term memory network, and the input information is a fusion feature set of N consecutive time units, used to predict the deformation trend within the next M minutes; the system's spatial analysis branch uses a convolutional neural network to derive crack shape, structural texture, etc. from the data generated by the above-mentioned three-dimensional rock mass model.

5. When establishing a time series prediction model according to the structure plan determined by the above clauses, the time series prediction is established using a two-layer long short-term memory network. The input layer combines the time data structure into one layer, and the output of the last layer combines the representations of each terminal into one layer. The dropout module is used in the middle to reduce sparsity. The convolutional neurons are composed of 3x3 convolution kernels and compound operations, and the speed is improved by reducing the dimensionality.

6. Based on the functional architecture defined in claims 3-5, the output tensors of the temporal and spatial analysis modules have the same dimension, and the channel dimension data will be interacted in the next step of feature integration and classification node implementation.

7. Based on the technical solution described in claim 6, the core part of the feature fusion and decision layer includes a fully connected layer and a classification output layer, wherein the number of neurons in the classification output layer corresponds to the risk level items in the rock mass stability assessment system.

8. According to the system described in claim 1, the dynamic threshold adjustment unit of the system works as follows: different benchmark threshold pairs corresponding to various working conditions are stored in advance; the benchmark threshold pairs that match the real-time working conditions data given by the "working condition information acquisition and input" module are searched; and a coefficient range between 0.7 and 1.3 is determined as a standard means of adjusting the benchmark threshold by referring to the instantaneous changes in the multi-dimensional fusion feature set, thereby forming a dynamically adjusted warning threshold.

9. A method for advanced support of tunnel rock mass stability, characterized in that, include: S1. Control the lidar and millimeter-wave radar installed on the same mobile platform to simultaneously detect the excavated rock mass in the tunnel and acquire high-precision point cloud data and radar echo data respectively. S2. After time-stamp synchronization and spatial coordinate calibration of the high-precision point cloud data and the radar echo data, they are fused to generate fused three-dimensional rock mass features. S3. Extract normalized multi-dimensional fusion features from the fused three-dimensional rock mass model. The multi-dimensional fusion features include: micro-deformation of the rock mass surface, accuracy of fracture edge contour and surface texture change rate obtained based on lidar data, and macro-displacement, deformation rate and echo energy change rate of the rock mass obtained based on millimeter-wave radar data. S4. Real-time acquisition or reception of current tunnel construction status information, wherein the status information includes at least preset geological parameters characterizing the rock mass properties and identification information of the currently implemented construction method. S5. Based on the multi-dimensional fusion feature set of the time series, predict the rock mass deformation trend within a preset time period in the future; S6. Dynamically correct the early warning threshold based on the working condition information provided by the working condition information acquisition and input module; S7. Based on the deformation trend prediction results and the dynamically corrected early warning threshold, generate the rock mass stability risk level; S8. Based on the rock mass stability risk level, match and output the corresponding support scheme from the preset support scheme library; S9. Execute the corresponding local audible and visual alarm and remote information push according to the rock mass stability risk level, and visualize the fused three-dimensional rock mass model and the support scheme.

Images