Intelligent evaluation and dynamic management system for tunnel construction risk based on deep learning

Abstract

The present application relates to the technical field of karst tunnel intelligent construction and management, in particular to a tunnel construction risk intelligent evaluation and dynamic management and control system based on deep learning, comprising: a multi-source data acquisition module that synchronously acquires point clouds, aerial images and time series monitoring data and pre-processes them; a multi-modal multi-scale fusion module that extracts multi-modal data features with different sub-networks and performs dynamic weighted fusion; a real-time modeling evaluation module that constructs a real-time three-dimensional geological model, calculates a risk level and generates a visual interface and a report; and a dynamic management and control module that triggers an early warning and pushes a management and control instruction according to the risk level and returns data to form a closed-loop management and control. The present application dynamically optimizes feature weights and scale ratios through multi-source data fusion and real-time evaluation mechanisms, and improves the timeliness and accuracy of risk evaluation. At the same time, through a three-dimensional visual model and a closed-loop management and control, risk is intuitively presented and strategies are automatically adjusted, significantly enhancing the response efficiency and decision reliability of construction safety.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction and management technology for karst tunnels, specifically a tunnel construction risk intelligent assessment and dynamic control system based on deep learning. Background Technology

[0002] As tunnel engineering progresses towards deeper burial and more complex geological conditions, construction safety risks are becoming increasingly prominent. Traditional risk management relies on manual inspections and single-source data monitoring, which struggles to comprehensively capture hidden risk factors such as dynamic changes in geological structures and abnormal stress in support structures. Furthermore, tunnel construction involves multi-source heterogeneous data, but existing technologies generally suffer from data silos, lacking the ability to collaboratively analyze and fuse features from multimodal data. In addition, the construction environment exhibits significant spatiotemporal dynamics, requiring risk assessment to consider both high-frequency real-time monitoring and low-frequency geological evolution characteristics. Traditional methods struggle to achieve dynamic coupling of multi-scale features. These technical bottlenecks lead to delayed risk warnings and insufficiently targeted control measures, becoming key challenges restricting the improvement of tunnel construction safety levels.

[0003] In summary, traditional tunnel risk management technologies have the following shortcomings:

[0004] 1. Data utilization is fragmented, relying solely on a single type of sensor or manual inspection records, leading to one-sided risk identification.

[0005] 2. The feature extraction method is crude, using fixed weights to fuse multi-source data, ignoring the differences in contribution of each modal data under different construction stages, and failing to establish a feature timeliness evaluation mechanism, making it difficult to adapt to scenarios with sudden changes in geological conditions.

[0006] 3. The control process is open-loop, and there is a lack of feedback on the effect after the early warning instructions are executed. It is impossible to adjust the control strategy according to the actual situation on site, which can easily lead to over-intervention or under-response. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a deep learning-based intelligent assessment and dynamic management system for tunnel construction risks. This system effectively overcomes the shortcomings of existing tunnel risk management technologies, such as one-sided risk identification, difficulty in adapting to sudden changes in geological conditions, and lack of feedback mechanisms after the execution of early warning commands.

[0008] To achieve the above objectives, the deep learning-based intelligent assessment and dynamic management system for tunnel construction risks of this invention mainly includes:

[0009] Multi-source data acquisition module: configured to acquire multi-dimensional heterogeneous data of the tunnel construction environment, including point cloud data, image data, and time-series monitoring data, and to preprocess the multi-dimensional heterogeneous data;

[0010] Multimodal multiscale fusion module: connected to the multi-source data acquisition module, configured to use a deep learning sub-network to extract basic feature vectors from point cloud, image and time-series monitoring data respectively; calculate the weights of each modality feature through modal prototype attention fusion algorithm and fuse them to obtain multimodal fusion features; split the multimodal fusion features according to sampling frequency, and then fuse them through the weights of each scale feature through time-varying feature scale attention algorithm to finally obtain multiscale fusion features;

[0011] Real-time modeling and evaluation module: connected to the multimodal and multiscale fusion module, configured to construct a real-time three-dimensional geological model based on the Poisson reconstruction algorithm, combining the point cloud data with multi-scale fusion features, and mapping the risk level information to the real-time three-dimensional geological model for visualization evaluation;

[0012] Dynamic control module: Connected to the real-time modeling and evaluation module, configured to automatically trigger graded early warnings based on risk assessment results, generate targeted control instructions based on a preset control strategy library and push them to the construction terminal, while simultaneously linking the access control system to execute control actions, and sending the feedback data after control back to the real-time modeling and evaluation module to form a closed-loop control.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] This invention, by deploying a 3D laser scanner, drones, and sensors, achieves simultaneous acquisition and preprocessing of point cloud data, aerial imagery data, and time-series monitoring data, providing multi-dimensional data support for risk assessment. The multimodal, multi-scale fusion module uses a sub-network architecture to extract the basic features of point cloud, imagery, and sensor data respectively, and dynamically calculates the weights of each modality feature through a modal prototype attention algorithm, effectively solving the problems of single data source and one-sided feature extraction in traditional methods. At the same time, the introduction of a time-varying feature scale attention algorithm enables the system to dynamically adjust the fusion ratio of high-frequency and low-frequency features according to the construction stage, significantly improving the timeliness and accuracy of risk assessment and providing a more reliable decision-making basis for tunnel construction safety.

[0015] This invention generates a 3D geological model through a real-time modeling and evaluation module, overlaying risk levels and anomaly information onto the model using color-coded markers. This visualizes the risks and significantly enhances the risk perception capabilities of construction personnel. The dynamic control module automatically triggers audible and visual alerts, SMS messages, and pop-up warnings based on preset risk thresholds. Simultaneously, it coordinates with the access control system to execute control actions and generates targeted instructions from a control strategy library, pushing them to the construction terminals. Finally, the feedback data after control is transmitted back to the real-time modeling and evaluation module, forming a closed-loop control system. This invention effectively avoids the problems of delayed instruction execution and passive risk response in traditional control methods. It not only significantly improves the response speed for construction safety but also enhances the intuitiveness and operability of risk control through 3D visualization technology. Attached Figure Description

[0016] Figure 1 This is a flowchart of the process of the present invention;

[0017] Figure 2 This is a diagram of the module data interaction framework of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 and Figure 2 As shown, the intelligent assessment and dynamic management system for tunnel construction risks based on deep learning includes: a multi-source data acquisition module, a multi-modal and multi-scale fusion module, a real-time modeling and evaluation module, and a dynamic management module.

[0020] The multi-source data acquisition module is configured to acquire multi-dimensional heterogeneous data of the tunnel construction environment. Specifically, it includes: a 3D laser scanning unit deployed in a stable area 5-10m behind the tunnel face to collect point cloud data of the 3D coordinates, surface texture, and crack distribution of the tunnel face and surrounding rock within this area; a drone aerial photography unit with 3-5 fixed take-off and landing points set up within 300-500m of the tunnel entrance to collect high-definition optical images and infrared thermal imaging data of the tunnel entrance, slope, and temporary support structures; and contact sensing units deployed with vibrating wire stress sensors, fiber optic displacement gauges, and submersible pressure sensors in steel supports, surrounding rock surfaces, and groundwater-rich areas to collect real-time time-series monitoring data of steel support stress, surrounding rock displacement, and groundwater pressure. These three types of data achieve full coverage of the tunnel construction scene characteristics from three dimensions: spatial 3D, macroscopic region, and temporal dynamics, compensating for the limitations of single data in feature representation.

[0021] In addition, this module is also used to perform format conversion, noise reduction, calibration, outlier removal, missing value completion and normalization on the collected raw data to form a standardized multi-source dataset for subsequent modules to call.

[0022] The multimodal multiscale fusion module is configured to perform the following steps:

[0023] 1. For point cloud data, PointNet++, with its hierarchical sampling and local neighborhood feature aggregation mechanism, perfectly adapts to its unstructured spatial characteristics. For image data, ResNet50 effectively solves the gradient vanishing problem in deep network training through its residual block structure, enabling efficient extraction of multi-dimensional visual features from shallow textures to deep structural defects. For sensor time-series data, Bi-LSTM fully leverages its strong time-series dependency modeling capability, accurately capturing temporal correlation features while exhibiting excellent robustness to noise and data gaps in the complex environment of tunnel construction. Therefore, PointNet++, ResNet50, and Bi-LSTM sub-networks are used to extract the basic feature vectors of point cloud, image, and sensor data, respectively. The specific structures of the three networks are as follows:

[0024] PointNet++ sub-network: The input layer receives preprocessed point cloud data; the sampling layer selects 40% of the key sampling points; the grouping layer performs sphere lookup with a radius of 0.3m; the feature extraction part contains 3 consecutive feature aggregation units, each of which consists of "1×1 convolutional layer, BN layer, LeakyReLU activation function, 3×3 convolutional layer, BN layer, LeakyReLU activation function, 1×1 convolutional layer" in sequence, and outputs a 256-dimensional feature vector.

[0025] The ResNet50 sub-network removes the fully connected layers of the original network, retaining four residual block groups (Conv2-x to Conv5-x). Conv2-x contains three residual blocks, each consisting of a 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer. Conv3-x to Conv5-x contain four, six, and three residual blocks, respectively, with the number of convolutional kernels in each residual block being 128 / 128 / 512, 256 / 256 / 1024, and 512 / 512 / 2048, respectively. A global average pooling layer and a 1×1 convolutional layer are added after Conv5-x, ultimately outputting a 256-dimensional feature vector.

[0026] Bi-LSTM subnetwork: The input layer receives standardized sensor time-series data, which is then mapped to a 128-dimensional vector by the embedding layer; the hidden layer consists of two bidirectional LSTM layers, with the first layer containing 128 neurons and the second layer containing 256 neurons. Each layer uses the Tanh activation function and has a dropout rate of 0.2; the output layer is a fully connected layer using the Linear activation function, and finally outputs a 256-dimensional feature vector.

[0027] 2. The modal prototype attention fusion algorithm is used to calculate the weights of each modality feature and then fuse them to obtain multimodal fused features; the calculation formula of the modal prototype attention fusion algorithm is as follows:

[0028] ,

[0029] in, It is a multimodal fusion feature, obtained by weighted fusion of basic feature vectors from point cloud, image, and sensor data. It is a modal identifier. These correspond to point cloud, image, and sensor data, respectively. It corresponds to the mode. The basic feature vector, For the corresponding mode The feature weights are calculated using the following formula:

[0030] ,

[0031] in, For risk sensitivity coefficient, This is a real-time data reliability index, calculated based on the error of the data acquisition equipment and the degree of environmental interference, with a value ranging from 0 to 1. The current modal feature vector With global feature prototype cosine similarity, The variance of the original data for each modality. It is the modal identifier in the summation term, and... The meaning is the same. These correspond to three modalities: point cloud, image, and sensor data, respectively. It corresponds to the mode. Risk sensitivity coefficient, It corresponds to the mode. Real-time data reliability index The current mode eigenvectors With global feature prototype cosine similarity, It corresponds to the mode. The variance of the original data.

[0032] 3. The multimodal fusion features are split according to the sampling frequency, and the weights of each scale feature are fused by the time-varying feature scale attention algorithm to finally obtain the multi-scale fusion features, which are then transmitted to the real-time modeling and evaluation module.

[0033] The calculation formula for the time-varying feature-scale attention algorithm is as follows:

[0034] ,

[0035] in, It is a multi-scale fusion feature, obtained by weighted fusion of high-frequency and low-frequency features. Corresponding to high-frequency features and low-frequency features respectively. It is the corresponding scale eigenvectors, The feature weights corresponding to scale t are calculated using the following formula:

[0036] ,

[0037] in, For time scale factors, high-frequency features Values ​​range from 0.6 to 0.9, indicating low-frequency characteristics. The value ranges from 0.1 to 0.4, and is dynamically adjusted according to the tunnel construction stage. The timeliness index is calculated from the mutual information between the feature vector and the risk label for the next hour. It is the characteristic scale identifier in the summation term. Corresponding to high-frequency and low-frequency characteristics respectively. It is the corresponding scale The time scale factor, It is the corresponding scale The characteristic timeliness index, It is the corresponding scale The degree of difference between current features and historical features and These represent the difference between high-frequency features and low-frequency features, respectively.

[0038] The real-time modeling and evaluation module is configured to perform the following steps:

[0039] 1. A real-time 3D geological model is constructed by using the Poisson reconstruction algorithm and combining geological structure information from point cloud data and multi-scale fusion features;

[0040] The calculation formula for the Poisson reconstruction algorithm is as follows:

[0041] ,

[0042] in, For the generated 3D mesh model, For the implicit signed distance function, The coordinates of a point in three-dimensional space. For function At point gradient at, For point cloud data in points The normal vector at that point, It is the variable that minimizes the objective function. Operators, It is a triple integral symbol, and the integration domain is . , For the three-dimensional reconstruction spatial domain, For the boundary of the spatial domain, Here are the boundary regularization coefficients. It is the symbol for a double integral, and the integration domain is . .

[0043] The system obtains the sign distance function by solving the variational equation. and by extracting The isosurfaces are used to generate a three-dimensional mesh model M.

[0044] The specific steps for constructing a real-time 3D geological model are as follows:

[0045] (1) Receive the preprocessed point cloud data transmitted by the multi-source data acquisition module and the geological structure information in the multi-scale fusion features transmitted by the multi-modal multi-scale fusion module. The point cloud data is in PCD format and contains three-dimensional coordinates, surface texture and fracture distribution information.

[0046] (2) Unify the coordinates of the received point cloud data and convert it to the tunnel construction design coordinate system to ensure that the spatial position of the point cloud data matches the design parameters;

[0047] (3) The Poisson reconstruction algorithm is adopted, using point cloud data as input, setting the reconstruction depth parameter to 10-12 and the sampling density parameter to 10-12, and generating an initial three-dimensional mesh model based on the spatial distribution characteristics of the point cloud data.

[0048] (4) Extract geological structure information from the multi-scale fusion features, including fracture density, rock layer strike and fault distribution data, map this information to the corresponding spatial location of the initial three-dimensional mesh model, and perform texture mapping and detail enhancement on the mesh model;

[0049] (5) Combine the tunnel cross-section dimensions and support structure parameters in the tunnel construction design CAD drawings to correct the boundary of the three-dimensional mesh model and ensure that the deviation between the model boundary and the design outline does not exceed 5cm;

[0050] (6) The modified three-dimensional model is lightened, the number of grids in non-critical areas is simplified, and the high-density grids in the face and surrounding 5m range are retained to generate the final real-time three-dimensional geological model.

[0051] (7) Repeat steps (1) to (6) above according to the same period as the laser scanner acquisition frequency to realize the real-time update of the three-dimensional geological model.

[0052] 2. Based on multi-scale fusion features, anomaly confidence, and geological condition parameters, a multi-dimensional risk coupling assessment algorithm is used to calculate the risk level;

[0053] The calculation formula for the multi-dimensional risk coupling assessment algorithm is as follows:

[0054] ,

[0055] in, It is a risk level. For the standardized function of fusion features, For historical data The maximum value, round This is a rounding function. Multi-scale fusion features This is the anomaly confidence level, with a value ranging from 0 to 1. This is the geological condition influence coefficient, outputting risk levels 1-6. or When, it is judged as low risk; when or When, it is judged as medium risk; when or At that time, it was determined to be high risk.

[0056] 3. The risk level and abnormal information are overlaid on the 3D model using color markings to generate a visualization interface and risk report, which are then transmitted to the dynamic management module;

[0057] The dynamic management module is configured to perform the following steps:

[0058] Based on the preset threshold of risk level, sound and light, SMS and pop-up warnings are activated, the control strategy library is called to generate targeted control instructions, which are pushed to the construction terminal, the data after control are collected to reassess the risk, and the warning is lifted or the strategy is adjusted according to the assessment results, forming a closed-loop control.

[0059] The system is pre-set with three risk levels, each corresponding to a differentiated hierarchical control strategy, as follows: Levels 1-2 (low risk) correspond to a blue alert, generating only a visual pop-up notification on the system management platform interface, without triggering on-site audible and visual alarms or SMS notifications; Levels 3-4 (medium risk) correspond to a yellow alert, immediately activating the on-site yellow audible and visual alarm, sending an alert SMS to the mobile terminals of the project safety manager and on-site construction team leader via the system communication interface, and simultaneously displaying a yellow top-level warning pop-up on the system management platform interface; Levels 5-6 (high risk) correspond to a red alert, immediately activating the on-site red audible and visual alarm, sending an emergency warning SMS to the mobile terminals of the project manager, safety director, supervising engineer, and on-site representative of the construction unit via the system communication interface, displaying a red full-screen mandatory warning pop-up on the system management platform interface, and simultaneously establishing a real-time communication link with the tunnel entrance access control system via the industrial control bus, issuing a locking execution command to the access control controller. Upon receiving the command, the access control system automatically cuts off personnel access authorization, locks the access gate, and only retains exclusive access permissions for authorized personnel such as project emergency command and rescue personnel, prohibiting unauthorized and irrelevant personnel from entering the construction area.

[0060] The system has a built-in control strategy library, which stores instruction templates based on a three-dimensional classification of "risk type - geological conditions - construction stage". The risk types include four categories: abnormal seepage pressure, sudden increase in displacement, stress exceeding limits, and environmental exceedances. The geological conditions are divided into three categories: soft rock, hard rock, and water-rich strata. The construction stages cover three stages: face excavation, initial support, and secondary lining. A total of 36 targeted instruction templates are formed. When an early warning is triggered, the system automatically matches the three-dimensional parameters corresponding to the current risk, calls the corresponding instruction template, and supplements real-time data.

[0061] Example 1: Abnormal seepage pressure scenario during tunnel face excavation in soft rock strata

[0062] A 3D laser scanner is deployed in a stable area 5-10m behind the tunnel face to collect 3D coordinates, surface texture, and point cloud data of crack distribution in the tunnel face and surrounding rock 5-10m behind it. Three fixed take-off and landing points are set up within 300-500m around the tunnel entrance to collect high-definition optical and infrared thermal aerial images of the tunnel entrance, slope, and temporary support structures using drones. Simultaneously, vibrating wire stress sensors are deployed at the steel support locations, fiber optic displacement gauges are deployed on the surrounding rock surface, and submersible pressure sensors are deployed in groundwater-rich areas to collect time-series monitoring data of steel support stress, surrounding rock displacement, and groundwater pressure. All collected data are processed by format conversion, noise reduction, calibration, outlier removal, missing value completion, and normalization before storage to provide high-quality data for subsequent modules.

[0063] The PointNet++, ResNet50, and Bi-LSTM sub-networks are used to extract basic feature vectors from point cloud, image, and sensor data, respectively. Specifically, the PointNet++ sub-network receives preprocessed point cloud data at its input layer, selects 40% of key sampling points at its sampling layer, and uses a grouping layer to perform sphere query to divide local regions with a radius of 0.3m. Features are then extracted through three consecutive feature aggregation units, ultimately outputting a 256-dimensional basic feature vector for the point cloud. The ResNet50 sub-network removes the original fully connected layers, retains four residual block groups, and adds a global average pooling layer and a 1×1 convolutional layer after Conv5-x, outputting a 256-dimensional basic feature vector for the image. The Bi-LSTM sub-network receives standardized sensor time-series data at its input layer, maps it to a 128-dimensional vector through an embedding layer, and outputs a 256-dimensional basic feature vector for the sensor through two bidirectional LSTM layers and a fully connected layer.

[0064] Subsequently, the modal prototype attention fusion algorithm is used to calculate the weights of each modality feature and then fuse them to obtain multimodal fused features. The calculation formula for the modal prototype attention fusion algorithm is as follows:

[0065] ,

[0066] in, It is a multimodal fusion feature, obtained by weighted fusion of basic feature vectors from point cloud, image, and sensor data. It corresponds to the mode. The basic feature vector, It is a modal identifier. These correspond to point cloud, image, and sensor data, respectively. For the corresponding mode The feature weights are calculated using the following formula:

[0067] ,

[0068] in, For risk sensitivity coefficient, This is a real-time data reliability index, calculated based on the error of the data acquisition equipment and the degree of environmental interference, with a value ranging from 0 to 1. The current modal feature vector With global feature prototype cosine similarity, The variance of the original data for each modality. It is the modal identifier in the summation term, and... The meaning is the same. These correspond to three modalities: point cloud, image, and sensor data, respectively. It corresponds to the mode. Risk sensitivity coefficient, It corresponds to the mode. Real-time data reliability index The current mode eigenvectors With global feature prototype cosine similarity, It corresponds to the mode. The variance of the original data.

[0069] The multimodal fusion features are then split according to the sampling frequency, and the multi-scale fusion features are calculated using a time-varying feature scale attention algorithm. The calculation formula for the time-varying feature scale attention algorithm is as follows:

[0070] ,

[0071] in, It is a multi-scale fusion feature, obtained by weighted fusion of high-frequency and low-frequency features. Corresponding to high-frequency features and low-frequency features respectively. It is the corresponding scale The eigenvectors of are calculated using the following formula:

[0072] ,

[0073] in, For time scale factors, high-frequency features Values ​​range from 0.6 to 0.9, indicating low-frequency characteristics. The value ranges from 0.1 to 0.4, and is dynamically adjusted according to the tunnel construction stage. The timeliness index is calculated from the mutual information between the feature vector and the risk label for the next hour. It is the characteristic scale identifier in the summation term. Corresponding to high-frequency and low-frequency characteristics respectively. It is the corresponding scale The time scale factor, It is the corresponding scale The characteristic timeliness index, It is the corresponding scale The difference between current features and historical features is transmitted to the real-time modeling and evaluation module.

[0074] In the real-time modeling and evaluation module, a real-time 3D geological model is first constructed. This involves receiving preprocessed PCD format point cloud data from the multi-source data acquisition module and geological structure information from the multi-scale fusion features transmitted by the multi-modal multi-scale fusion module. The received point cloud data is then coordinate-unified and transformed to the tunnel construction design coordinate system to ensure spatial alignment with design parameters. The Poisson reconstruction algorithm is then used, with the point cloud data as input. The reconstruction depth parameter is set to 10, and the sampling density parameter is set to 10. An initial 3D mesh model is generated based on the spatial distribution characteristics of the point cloud data. The calculation formula for the Poisson reconstruction algorithm is as follows:

[0075] ,

[0076] in, For the generated 3D mesh model, For the implicit signed distance function, The coordinates of a point in three-dimensional space. For function At point gradient at, For point cloud data in points The normal vector at that point, It is the variable that minimizes the objective function. Operators, It is a triple integral symbol, and the integration domain is . , For the three-dimensional reconstruction spatial domain, For the boundary of the spatial domain, Here are the boundary regularization coefficients. It is the symbol for a double integral, and the integration domain is . .

[0077] The sign distance function is obtained by solving this variational equation. Then by extraction The isosurfaces are used to generate a three-dimensional mesh model.

[0078] Subsequently, geological structural information such as fracture density, stratum strike, and fault distribution from the multi-scale fusion features were extracted and mapped to the corresponding spatial locations of the initial 3D mesh model for texture mapping and detail enhancement. Combined with the tunnel cross-section dimensions and support structure parameters in the tunnel construction design CAD drawings, the boundary of the 3D mesh model was corrected to ensure that the deviation between the model boundary and the design outline does not exceed 5cm. The corrected 3D model was then lightweighted, simplifying the number of meshes in non-critical areas while retaining high-density meshes within 5m of the tunnel face and surrounding area to generate the final real-time 3D geological model, which was then updated at a cycle consistent with the acquisition frequency of the laser scanner.

[0079] Next, a risk assessment is conducted. Based on multi-scale fusion features, anomaly confidence levels, and geological condition parameters, the risk level is calculated using a multi-dimensional risk coupling assessment algorithm. The calculation formula for the multi-dimensional risk coupling assessment algorithm is as follows:

[0080] ,

[0081] in, It is a risk level. For the standardized function of fusion features, For historical data The maximum value, round This is a rounding function. Multi-scale fusion features This is the anomaly confidence level, with a value ranging from 0 to 1. This is the geological condition influence coefficient, outputting risk levels 1-6. or When, it is judged as low risk; when or When, it is judged as medium risk; when or At that time, it was determined to be high risk. After calculation, due to the abnormal groundwater seepage pressure, the risk level was determined to be Level 5 high risk. Finally, the Level 5 risk level and seepage pressure abnormality information were superimposed on the 3D model through color marking to generate a visualization interface and risk report, which were then transmitted to the dynamic management module.

[0082] Based on preset thresholds, the dynamic control module activates a red alert due to a risk level of 5 (high risk). This includes activating a red audible and visual alarm on-site, sending text messages to the project manager, safety director, supervising engineer, and on-site representative of the construction unit, displaying a red full-screen pop-up on the system platform, and simultaneously linking the tunnel entrance access control system to prevent unauthorized personnel from entering the construction area.

[0083] Subsequently, the control strategy library is invoked to generate targeted control instructions. The control strategy library stores instruction templates categorized into three dimensions: risk type, geological conditions, and construction stage. First, parameter matching is performed. For risk type parameter matching, risk characteristics are extracted from the abnormal information output by the real-time modeling and evaluation module and compared with the preset seepage pressure anomaly characteristic parameters in the control strategy library to determine that the current risk type is seepage pressure anomaly. For geological condition parameter matching, lithology, water content, and rock mass integrity parameters contained in the three-dimensional geological model in the real-time modeling and evaluation module are read and matched with the soft rock strata division parameters in the control strategy library to determine that the current geological condition is soft rock strata. For construction stage parameter matching, the current tunnel construction procedure information recorded in the construction management system is obtained and compared with the face excavation procedure characteristics in the control strategy library to determine that the current construction stage is the face excavation stage.

[0084] Based on the matched three-dimensional parameters of "abnormal seepage pressure - soft rock strata - excavation face", the corresponding instruction template is located in the control strategy library. Real-time monitoring data related to the current abnormal seepage pressure risk is retrieved from the multi-source data acquisition module, and the risk location coordinates and 5-level risk level data are obtained from the real-time modeling and assessment module. These data are filled into the preset data placeholder positions of the instruction template to form a complete control instruction and push it to the construction terminal. After the construction personnel take corresponding measures according to the control instruction, the system collects the controlled data and sends it back to the multi-modal multi-scale fusion module to reassess the risk. If the risk is reduced, the warning is lifted; if the risk is still high, the strategy is adjusted to form a closed-loop control.

[0085] In summary, in the scenario of abnormal seepage pressure during the excavation stage of a tunnel face in soft rock strata, the system acquires high-quality data of various types through a multi-source data acquisition module, laying the foundation for subsequent stages; through a multi-modal and multi-scale fusion module, features are extracted using specific sub-networks and the two algorithms are fused to obtain comprehensive features; the real-time modeling and evaluation module constructs an accurate and real-time updated three-dimensional geological model, and calculates a level 5 high risk based on the algorithm; the dynamic control module activates a red alert, generates and pushes control instructions based on the three-dimensional parameters, and forms a closed-loop control through data feedback, effectively addressing the risk of abnormal seepage pressure and ensuring construction safety.

[0086] Example 2: Situation of sudden displacement increase during the initial support stage of a tunnel in hard rock strata

[0087] A 3D laser scanner is deployed in a stable area 5-10m behind the tunnel face to collect 3D coordinates, surface texture, and point cloud data of crack distribution in the tunnel face and surrounding rock 5-10m behind it. Five fixed take-off and landing points are set up 300-500m around the tunnel entrance, and high-definition optical and infrared thermal imaging aerial images of the tunnel entrance, slope, and temporary support structure are collected using drones. Vibrating wire stress sensors are deployed at the steel supports, fiber optic displacement gauges are deployed on the surrounding rock surface, and submersible pressure sensors are deployed in groundwater-rich areas to collect time-series monitoring data of steel support stress, surrounding rock displacement, and groundwater pressure. The collected data are then processed sequentially, including format conversion, noise reduction, calibration, outlier removal, missing value completion, and normalization, before being stored to ensure data availability.

[0088] The basic feature vectors of point cloud, image, and sensor data are extracted using PointNet++, ResNet50, and Bi-LSTM sub-networks, respectively. The PointNet++ sub-network takes the preprocessed point cloud data as input, selects 40% of the key sampling points through the sampling layer, and divides the local region by spherical query with a radius of 0.3m through the grouping layer. After processing by three feature aggregation units, it outputs a 256-dimensional basic feature vector of the point cloud. The ResNet50 sub-network retains four residual block groups, removes the fully connected layer, and adds a global average pooling layer and a 1×1 convolutional layer after Conv5-x to output a 256-dimensional basic feature vector of the image. The Bi-LSTM sub-network processes the standardized sensor time-series data, and outputs a 256-dimensional basic feature vector of the sensor through embedding layer mapping, bidirectional LSTM layer and fully connected layer.

[0089] The modal prototype attention fusion algorithm is used to calculate and fuse the feature weights of each modality to obtain multimodal fused features. The calculation formula of the modal prototype attention fusion algorithm is as follows:

[0090] ,

[0091] After splitting according to sampling frequency, multi-scale fusion features are calculated using a time-varying feature scale attention algorithm. The calculation formula for the time-varying feature scale attention algorithm is as follows: It is then transmitted to the real-time modeling and evaluation module.

[0092] When constructing a real-time 3D geological model, preprocessed PCD format point cloud data and geological structure information from multi-scale fusion features are received; coordinates are unified to the tunnel construction design coordinate system; the Poisson reconstruction algorithm is used, with reconstruction depth parameter 12 and sampling density parameter 12 set, to generate an initial 3D mesh model from the point cloud data. The calculation formula for the Poisson reconstruction algorithm is as follows:

[0093] ,

[0094] The sign distance function is obtained by solving this variational equation. Then by extraction The isosurface is used to generate a 3D mesh model; the geological structure information in the multi-scale fusion features is mapped to the initial model for texture mapping and detail enhancement; the model boundary is corrected in conjunction with CAD drawings to ensure that the deviation does not exceed 5cm; the model is lightweighted, retaining the 5m high-density mesh of the working face and surrounding area, generating a real-time 3D geological model and updating it according to the laser scanner acquisition frequency.

[0095] When conducting risk assessment, a multi-dimensional risk coupling assessment algorithm is used to calculate the risk level based on multi-scale fusion features, anomaly confidence levels, and geological condition parameters. The calculation formula for the multi-dimensional risk coupling assessment algorithm is as follows:

[0096] ,

[0097] Output risk levels 1-6, when or When, it is judged as low risk; when or When, it is judged as medium risk; when or At that time, it was judged as high risk. This time, due to the sudden increase in surrounding rock displacement, the risk level was judged as level 4 medium risk. The level 4 risk level and the information on the sudden increase in displacement were superimposed on the 3D model through color marking to generate a visualization interface and risk report, which were then transmitted to the dynamic control module.

[0098] The dynamic control module has activated a yellow alert due to its risk level of 4 (medium risk). This means that the yellow audible and visual alarm on site is activated, and text messages are sent to the project safety manager and the on-site construction team leader. A yellow pop-up window appears on the system platform.

[0099] The control strategy library is invoked to generate control instructions. First, parameter matching is performed. In terms of risk type, the risk characteristics in the abnormal information are extracted and compared with the displacement surge characteristic parameters to determine that the risk type is displacement surge. In terms of geological conditions, the lithology, water content, and rock mass integrity parameters in the three-dimensional geological model are read and matched with the hard rock strata division parameters to determine that the geological conditions are hard rock strata. In terms of construction stage, the construction procedure information is obtained and compared with the initial support procedure characteristics to determine that the construction stage is the initial support stage.

[0100] Based on the three-dimensional parameters of "displacement surge - hard rock strata - initial support", the corresponding instruction template is found in the control strategy library. Relevant real-time monitoring data, risk location coordinates and level 4 risk data are retrieved and filled into the template data placeholders to form a complete control instruction that is pushed to the construction terminal. After the construction is completed, the control data is collected and transmitted back to reassess the risk. Based on the assessment results, the warning is lifted or the strategy is adjusted to achieve closed-loop control.

[0101] In summary, for the scenario of sudden displacement increase during the initial support stage of tunnels in hard rock strata, the multi-source data acquisition module collects and processes various types of data; the multi-modal multi-scale fusion module extracts features using corresponding sub-networks and obtains multi-scale fused features through two algorithms; the real-time modeling and evaluation module constructs a compliant 3D geological model and determines the risk level as level 4 (medium risk) based on algorithms; the dynamic control module activates a yellow alert, matches instruction templates based on 3D dimension parameters, supplements data to generate control instructions, and combines data feedback to achieve closed-loop control, helping to solve the problem of sudden displacement increase and maintain construction stability.

[0102] The embodiments of the present invention have been described in detail above, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A deep learning-based intelligent risk assessment and dynamic control system for tunnel construction, characterized in that, include: Multi-source data acquisition module: configured to acquire multi-dimensional heterogeneous data of the tunnel construction environment, including point cloud data, image data, and time-series monitoring data, and to preprocess the multi-dimensional heterogeneous data; Multimodal multiscale fusion module: connected to the multi-source data acquisition module, configured to extract basic feature vectors from point cloud, image and time-series monitoring data respectively using deep learning sub-network; calculate the weights of each modality feature through modal prototype attention fusion algorithm and fuse them to obtain multimodal fusion features; split the multimodal fusion features according to sampling frequency, and then calculate the weights of each scale feature through time-varying feature scale attention algorithm and fuse them to finally obtain multiscale fusion features; Real-time modeling and evaluation module: connected to the multimodal and multiscale fusion module, configured to construct a real-time three-dimensional geological model based on the Poisson reconstruction algorithm, combining the point cloud data with multi-scale fusion features, and mapping the risk level information to the real-time three-dimensional geological model for visualization evaluation; Dynamic control module: Connected to the real-time modeling and evaluation module, configured to automatically trigger graded early warnings based on risk assessment results, generate targeted control instructions based on a preset control strategy library and push them to the construction terminal, while simultaneously linking the access control system to execute control actions, and sending the feedback data after control back to the real-time modeling and evaluation module to form a closed-loop control.

2. The intelligent assessment and dynamic control system for tunnel construction risks based on deep learning according to claim 1, characterized in that, The deep learning sub-networks include the PointNet++ sub-network, the ResNet50 sub-network, and the Bi-LSTM sub-network.

3. The intelligent assessment and dynamic control system for tunnel construction risks based on deep learning according to claim 1, characterized in that, The calculation formula for the modal prototype attention fusion algorithm in the multimodal multiscale fusion module is as follows: ， in, It is a multimodal fusion feature. It is a modal identifier. These correspond to point cloud, image, and sensor data, respectively. It corresponds to the mode. The basic feature vector, For the corresponding mode The feature weights are calculated using the following formula: ， in, For risk sensitivity coefficient, This is a real-time data reliability index, with a value ranging from 0 to 1. The current modal feature vector With global feature prototype cosine similarity, The variance of the original data for each modality. It is the modal identifier in the summation term, and... The meaning is the same. These correspond to three modalities: point cloud, image, and sensor data, respectively. It corresponds to the mode. Risk sensitivity coefficient, It corresponds to the mode. Real-time data reliability index The current mode eigenvectors With global feature prototype cosine similarity, It corresponds to the mode. The variance of the original data.

4. The intelligent assessment and dynamic control system for tunnel construction risks based on deep learning according to claim 1, characterized in that, The calculation formula for the time-varying feature scale attention algorithm in the multimodal multiscale fusion module is as follows: ， in, It is a multi-scale fusion feature. Corresponding to high-frequency features and low-frequency features respectively. It is the corresponding scale eigenvectors, The feature weights corresponding to scale t are calculated using the following formula: ， in, For time scale factors, high-frequency features Values ​​range from 0.6 to 0.9, indicating low-frequency characteristics. The value ranges from 0.1 to 0.

4. As a characteristic timeliness index, It is the characteristic scale identifier in the summation term. Corresponding to high-frequency and low-frequency characteristics respectively. It is the corresponding scale The time scale factor, It is the corresponding scale The characteristic timeliness index, It is the corresponding scale The degree of difference between current features and historical features This represents the difference between the current feature and historical features at the corresponding scale t. and These represent the difference between high-frequency features and low-frequency features, respectively.

5. The intelligent assessment and dynamic control system for tunnel construction risks based on deep learning according to claim 1, characterized in that, In the real-time modeling and evaluation module, the calculation formula for the Poisson reconstruction algorithm is as follows: ， in, For the generated 3D mesh model, For the implicit signed distance function, The coordinates of a point in three-dimensional space. For function At point gradient at, For point cloud data in points The normal vector at that point, It is the variable that minimizes the objective function. Operators, It is a triple integral symbol, and the integration domain is . , For the three-dimensional reconstruction spatial domain, For the boundary of the spatial domain, Here are the boundary regularization coefficients. It is the symbol for a double integral, and the integration domain is . .

6. The intelligent assessment and dynamic control system for tunnel construction risks based on deep learning according to claim 1, characterized in that, The specific steps for constructing a real-time 3D geological model in the real-time modeling and evaluation module are as follows: (1) Receive preprocessed point cloud data from the multi-source data acquisition module and geological structure information from the multi-scale fusion features of the multi-modal multi-scale fusion module; (2) Convert the point cloud data to the tunnel construction design coordinate system using coordinate system one; (3) Using the Poisson reconstruction algorithm, an initial three-dimensional mesh model is generated based on the spatial distribution characteristics of point cloud data; (4) Extract the geological structure information from the multi-scale fusion features and map it to the corresponding spatial location of the initial three-dimensional mesh model; (5) Based on the parameters in the CAD drawings of the tunnel construction design, the boundary of the three-dimensional mesh model is corrected so that the deviation between the model boundary and the design outline is controlled within 5cm. (6) Lightweight processing is performed on the corrected three-dimensional model to simplify the number of grids in non-critical areas, while retaining the high-density grids within 5m of the working face and surrounding area to generate the final real-time three-dimensional geological model. (7) Repeat steps (1) to (6) according to the same cycle as the laser scanner acquisition frequency to achieve real-time updating of the three-dimensional geological model.

7. The intelligent assessment and dynamic control system for tunnel construction risks based on deep learning according to claim 1, characterized in that, The risk level information in the real-time modeling and assessment module is calculated by a multi-dimensional risk coupling assessment algorithm, and the formula is as follows: ， in, It is a risk level. For the standardized function of fusion features, For historical data The maximum value, round This is a rounding function. Multi-scale fusion features This is the anomaly confidence level, with a value ranging from 0 to 1. This is the geological condition influence coefficient, outputting risk levels 1-6. or When, it is judged as low risk; when or When, it is judged as medium risk; when or At that time, it was determined to be high risk.

8. The intelligent assessment and dynamic control system for tunnel construction risks based on deep learning according to claim 1, characterized in that, The tiered early warning system in the dynamic control module specifically includes: (1) Blue alert: corresponds to level 1-2 low risk, generates a visual pop-up prompt on the system management platform, and does not trigger sound, light and SMS notifications; (2) Yellow warning: corresponding to level 3-4 medium risk, activate the yellow sound and light alarm on site, send warning text messages to the mobile terminals of the project safety department supervisor and the on-site construction team leader through the communication interface, and generate a yellow pop-up window on the system management platform; (3) Red Alert: For level 5-6 high risk, activate the red sound and light alarm on site, send emergency text messages to the mobile terminals of the project manager, safety director, supervising engineer and on-site representative of the construction unit through the communication interface, generate a red full-screen warning pop-up on the system management platform, and send a lock command to the tunnel entrance access control system to prohibit irrelevant personnel from entering the construction area.

9. The intelligent assessment and dynamic control system for tunnel construction risks based on deep learning according to claim 1, characterized in that, The generation of control commands in the dynamic control module relies on a pre-set control strategy library. The strategy library uses a three-dimensional matrix architecture of "risk type-geological conditions-construction stage" to classify and store command templates. The risk types include four categories: abnormal seepage pressure, sudden increase in displacement, stress exceeding limits, and environmental exceedances. The geological conditions are divided into three categories: soft rock, hard rock, and water-rich strata. The construction stages include three stages: face excavation, initial support, and secondary lining. When the system triggers an early warning, it automatically extracts the three-dimensional parameters under the current working condition and performs matching indexing in the strategy library, calls the corresponding command template, and fills the command template with real-time monitoring data to generate the final control command.

Images