Multi-source information fusion tunnel geological disaster intelligent prediction method and system

Abstract

The application discloses a multi-source information fusion tunnel geological disaster intelligent prediction method and system, belongs to the field of tunnel and underground engineering construction safety technology, and comprises four steps of multi-source heterogeneous data acquisition and structured conversion, dynamic risk grading and verification strategy matching, multi-level multi-source information fusion and intelligent identification, and prediction result output and feedback optimization. The application dynamically adjusts the weight of each data source through a self-adaptive confidence weighted fusion algorithm, adopts a fusion identification model combining a multi-branch deep neural network and a dynamic Bayesian network to output a comprehensive risk index, and realizes online incremental learning through prediction-revelation error feedback. Compared with the prior art, the application significantly weakens the multi-solution property of a single method, systematically incorporates a while-drilling parameter into a geological identification system for the first time, constructs a new intelligent dynamic prediction paradigm driven by risk grading, and realizes continuous evolution of prediction capability.

Description

Technical Field

[0001] This invention belongs to the field of tunnel and underground engineering construction safety technology, specifically involving a method and system that integrates multiple detection and sensing technologies, and uses multi-source information fusion and artificial intelligence methods to intelligently predict, dynamically assess and classify geological hazards ahead of the tunnel face during tunnel construction. Background Technology

[0002] In tunnel construction, accurate and timely prediction of hidden geological hazards ahead of the tunnel face is a core element in ensuring construction safety, controlling engineering risks, and optimizing construction decisions. Currently, advanced geological forecasting for tunnel construction mainly relies on geophysical exploration and advanced drilling, but the existing technological system has the following significant limitations, constituting key bottlenecks restricting the industry's safety and efficiency.

[0003] First, existing technologies face a significant contradiction between ambiguity and reliability. Currently, commonly used geophysical methods in engineering, such as seismic wave methods, ground-penetrating radar, and transient electromagnetic methods, inherently possess ambiguity in their interpretation results. That is, the same geophysical response may correspond to multiple different geological conditions. This leads to insufficient accuracy in spatial location, scale assessment, and property identification of key adverse geological bodies such as faults and water-rich bodies, resulting in both false alarms and missed alarms. Furthermore, the lack of effective data fusion and cross-validation mechanisms among these methods makes it difficult to form comprehensive and reliable deterministic judgments.

[0004] Secondly, obtaining in-situ mechanical information is difficult, and the information sources are not fully utilized. Advanced geological forecasting focuses on detecting anomalies in the physical field, lacking the ability to continuously acquire the actual mechanical parameters of the rock mass in the field. Although the thrust, torque, and rotational speed generated during drilling or excavation by a rock drilling rig reflect the degree of fragmentation, hardness, and integrity of the rock mass ahead in real time and continuously, and are valuable sources of in-situ sensing information, this data is currently mainly used for equipment status monitoring and has not yet been systematically processed, analyzed, and integrated into the geological identification system, resulting in a serious waste of information resources.

[0005] Furthermore, the existing forecasting system suffers from information silos and unstructured data. Forecasting work is carried out in stages by different units using different equipment, and the results are mostly delivered in the form of unstructured reports and drawings. The core conclusions, abnormal mileages, and characteristic parameters are locked in documents and cannot be directly read by computers or used for subsequent quantitative comprehensive analysis, historical data mining, and model training, thus forming information silos.

[0006] Furthermore, the forecasting system lacks hierarchy and dynamism. Existing forecasting models are usually a fixed combination of long-distance geophysical exploration and a small amount of drilling verification. They lack a decision-making mechanism that can dynamically and intelligently match and activate short-distance verification methods of different accuracies and costs based on the risk level of the preliminary forecast. This has failed to build a progressive and closed-loop forecasting system, resulting in suboptimal resource allocation, insufficient verification in high-risk areas, and excessive exploration in low-risk areas.

[0007] Finally, the identification process relies heavily on personal experience and has a low level of intelligence. The final geological interpretation and risk assessment depend heavily on the personal experience of geological engineers, with inconsistent identification standards that are difficult to pass on and scale up. Currently, the application of artificial intelligence and machine learning methods for the automatic fusion and intelligent identification of multi-source forecast data is still in the research and exploration stage, lacking a set of engineering-based and systematic solutions. Therefore, there is an urgent need for a comprehensive geological hazard prediction technology for tunnel construction that can integrate multi-source heterogeneous data, achieve complementary information advantages, possess intelligent identification and dynamic decision-making capabilities, and continuously self-optimize. Summary of the Invention

[0008] This invention aims to address key technical challenges in existing tunnel advanced geological prediction technologies, such as significant ambiguity, insufficient information utilization, rigid systems, and low levels of intelligence. The primary objective of this invention is to provide an intelligent prediction method and system for tunnel geological hazards based on multi-source information fusion and machine learning. This system achieves a transformation from single-method interpretation to multi-source fusion identification, from static prediction to dynamic evaluation, and from experience-driven to data- and model-driven approaches, ultimately significantly improving the accuracy, reliability, and real-time decision support of geological hazard prediction.

[0009] To achieve the above objectives, this invention provides a multi-source information fusion method for intelligent prediction of tunnel geological hazards. Its core lies in constructing a closed-loop intelligent decision-making system that integrates long-distance forecasting, risk classification, short-distance targeted verification, multi-source fusion identification, and real-time feedback optimization.

[0010] Step S1 involves the acquisition and structured transformation of multi-source heterogeneous data. This process systematically collects multi-source heterogeneous data generated during tunnel construction and converts it into a structured data list indexed by tunnel mileage, constructing a spatiotemporally aligned multi-source geological database. The multi-source heterogeneous data includes geophysical data, drilling parameter data, tunnel face 3D point cloud data, and borehole visual data. The geophysical report is intelligently analyzed to extract abnormal mileage intervals and physical parameters, generating a geophysical feature list. Drilling parameters are filtered and their features are calculated to extract the drilling specific energy index, generating a mechanical response list. The tunnel face point cloud is subjected to structural surface identification and attitude calculation, generating a structural surface feature list.

[0011] Step S2 is the dynamic risk grading and verification strategy matching step. Based on the multi-source data obtained in step S1, a quantitative risk assessment model is constructed to calculate the initial risk index of the segment to be predicted. The risk level is determined by comparing the initial risk index with a preset risk level threshold. Based on the risk level, the corresponding differentiated short-range verification strategy is automatically triggered by the verification strategy matching engine. Specifically, a low-risk level triggers a conventional verification strategy, a medium-risk level triggers an enhanced verification strategy, and a high-risk level triggers the highest-level verification strategy.

[0012] Step S3 involves multi-level, multi-source information fusion and intelligent identification. The long-range forecast features and short-range verification features obtained in steps S1 and S2 are spatiotemporally aligned and normalized in a unified mileage coordinate system to construct a multi-dimensional feature matrix. An adaptive confidence-weighted fusion algorithm is used to calculate the fusion weights of each data source, which are dynamically adjusted based on the degree of evidence conflict between the data sources. The multi-dimensional feature matrix is ​​input into the fusion identification model, which outputs the probability of occurrence and comprehensive risk index of various geological hazards.

[0013] Step S4 involves outputting and optimizing the prediction results, classifying risk levels and issuing early warnings based on the comprehensive risk index output in step S3. The actual geological conditions revealed by the excavation are used as ground truth labels and compared with the prediction results to calculate the prediction error. Based on the prediction error, the fusion identification model undergoes online incremental learning and adaptive parameter optimization to achieve continuous evolution of prediction capabilities.

[0014] This invention also provides a multi-source information fusion intelligent prediction system for tunnel geological hazards, implementing the above-mentioned methods. The system includes a data acquisition module, a risk classification module, a fusion identification module, and a feedback optimization module. The data acquisition module systematically collects multi-source heterogeneous data and constructs a spatiotemporally aligned multi-source geological database. The risk classification module calculates the initial risk index, determines the risk level, and triggers a differentiated verification strategy. The fusion identification module performs multi-source information fusion and intelligent identification. The feedback optimization module is used for early warning issuance and online incremental learning optimization of the model.

[0015] Compared with existing technologies, this invention has the following advantages: the forecast accuracy is significantly improved, and the ambiguity of a single method is fundamentally weakened through cross-validation and deep fusion of multi-source information; the depth and breadth of information utilization are unprecedented, and for the first time, drilling parameters are systematically incorporated into the geological identification system; a new paradigm of intelligent dynamic forecasting is constructed, and a dynamic decision-making mechanism of risk classification and verification matching is innovatively proposed; the identification process is made intelligent and standardized; and it has continuous evolution capability, enabling the forecast model to continuously adapt and optimize as the project progresses through a closed-loop feedback mechanism. Attached Figure Description

[0016] Figure 1This is an overall flowchart of the method of the present invention.

[0017] Figure 2 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment takes a high-speed railway tunnel project under construction as the application background to illustrate the complete implementation process of the present invention.

[0019] Step S1: Multi-source heterogeneous data acquisition and structured conversion. The core objective of this step is to systematically acquire multi-source heterogeneous data generated throughout the tunnel construction process and utilize intelligent sensing and processing technologies to uniformly convert it into a structured data list with tunnel mileage as the key index, thereby constructing a spatiotemporally aligned multi-source geological database. In one embodiment of this invention, the multi-source heterogeneous data includes four categories: geophysical data, drilling parameter data, tunnel face 3D point cloud data, and borehole visual data.

[0020] For intelligent parsing of geophysical exploration reports, this invention employs a combination of optical character recognition (OCR) and natural language processing (NLP) technologies to achieve automated processing. Specifically, the PDF file of the geophysical exploration report is first converted to image format using an image processing library. Then, the reflectance profile is binarized and its coordinate axes are identified. In this embodiment, a threshold segmentation method is used for binarization, with the threshold set to 0.65 times the grayscale value. The reflectance intensity curve is extracted through a pixel-mileage mapping relationship, where the pixel-mileage mapping coefficient is automatically calculated based on the image coordinate axis annotations. Preferably, a reflectance intensity threshold of 0.7 is set. When the reflectance intensity exceeds this threshold, it is identified as an abnormal segment, and the starting mileage, ending mileage, and peak intensity are recorded.

[0021] For the analysis of textual conclusions, this invention employs a pre-trained named entity recognition model for automatic extraction. In one embodiment of this invention, the named entity recognition model is obtained by fine-tuning a pre-trained language model on a self-built geological report corpus. This model can automatically extract three types of entities from textual conclusions: mileage intervals, anomaly features, and inferred conclusions. Preferably, the geological report corpus contains no fewer than 5,000 historical geophysical reports, covering descriptive texts of typical geological hazard types such as fault fracture zones, water-rich cavities, and weak interlayers. Taking this embodiment as an example, in the sentence segment DK158+220 to DK158+260, the reflected wave phase axis is discontinuous, and the wave velocity decreases by 20%, inferred to be a fault fracture zone. The model automatically extracts the mileage interval from 158220m to 158260m, the anomaly features as discontinuous phase axis and decreased wave velocity, and the inferred conclusion as a fault fracture zone. The final result is a structured list of geophysical features, with each record containing fields such as project number, data source, date, section number, starting mileage, ending mileage, feature type, feature value, and inferred risk type.

[0022] For feature extraction of drilling parameters, the goal of this step is to transform the drilling parameters of the drilling rig or tunnel boring machine into quantitative indicators reflecting the mechanical properties of the rock mass. In one embodiment of this invention, the original drilling parameters include feed pressure, slewing pressure, impact pressure, and drill pipe displacement, with a sampling frequency of 10Hz. First, data cleaning and calibration are performed. A Butterworth low-pass filter is used to remove high-frequency noise, with the filter cutoff frequency set to 2Hz and the filter order set to 4th. Gravity compensation is applied to the feed pressure based on the drill pipe inclination angle to calculate the actual thrust F, in kN. The torque T is obtained by converting the slewing pressure, in kN·m.

[0023] The Drilling Energy Index (SEI) is a key indicator comprehensively reflecting the drillability and strength of rock mass. In one embodiment of this invention, each 0.1m drilling depth is considered an analysis unit, and the characteristic parameters within that unit are calculated. The formula for calculating the SEI is as follows: , Among them, SEI is the drilling specific energy index, with the unit MJ / m³, which represents the energy consumed by the breaking of a unit volume of rock. The higher the value, the harder and more intact the rock mass. A sharp drop in the value indicates that there may be a weak or broken geological body ahead. F is the drilling thrust, with the unit kN, and the value range is 0 to 200 kN. It is acquired in real time by the drilling acquisition system and calibrated by gravity compensation. The penetration velocity is measured in m / min, ranging from 0 to 5 m / min, and is calculated from the time derivative of the drill pipe displacement signal; T is the drilling torque, measured in kN·m, ranging from 0 to 10 kN·m, and is calculated from the rotational pressure based on the hydraulic system characteristic curve; RPM is the drill pipe rotation speed, measured in r / min, ranging from 0 to 300 r / min, and is directly measured by the rotation speed sensor; A is the borehole cross-sectional area, measured in m², and is calculated based on the drill bit diameter. In this embodiment, the drill bit diameter is 76 mm, corresponding to a cross-sectional area of ​​0.00454 m².

[0024] Simultaneously, the thrust fluctuation coefficient is calculated as an evaluation index of rock mass homogeneity, and its formula is the thrust standard deviation divided by the thrust mean. When the fluctuation coefficient exceeds 0.3, it indicates that the rock mass may have significant heterogeneity or fracture development. Finally, a list of mechanical responses distributed along the borehole trajectory is generated, with each record containing fields such as mileage, average thrust, average torque, drilling specific energy index, and fluctuation coefficient.

[0025] For the 3D point cloud analysis of the tunnel face, this step uses 3D laser scanning technology to acquire high-precision geometric information of the tunnel face, and automatically identifies and quantifies the structural features of the rock mass through point cloud processing algorithms. In one embodiment of the present invention, the 3D laser scanner is installed at a sidewall position approximately 50m from the tunnel face, with a scanning accuracy better than 5mm, and a single scan yields approximately 20 million point clouds. Point cloud preprocessing is first performed, including statistical filtering for noise reduction and voxel downsampling, with the downsampled voxel size set to 10mm.

[0026] Structural facet identification employs a region-growing-based planar segmentation algorithm. Preferably, the normal vector deviation threshold for planar segmentation is set to 10°, and the minimum number of planar points is set to 1000. Principal component analysis is used to calculate the normal vector for each identified planar segment, which is then converted into geologically common strike, dip, and dip angle representations. In one embodiment of the invention, K-means clustering is used to group the attitude, with the number of clusters automatically determined based on the profile coefficient, typically 2 to 4 groups. The spacing and trace length of dominant joints in each group are statistically analyzed. The spacing is calculated using the vertical distance between adjacent parallel structural faces, and the trace length is estimated using the simulated window method. Simultaneously, the number of volumetric joints Jv in the rock mass is calculated, in units of joints / m³, to evaluate the integrity of the rock mass. Finally, a structural facet feature list is generated for the current working face, including the current mileage, rock mass quality score, number of dominant joint groups, attitude parameters of each joint group, spacing, extension, and information on the number of volumetric joints and weathering degree.

[0027] For borehole visual information analysis, when conducting advanced horizontal drilling, this invention configures a digital borehole camera system to acquire panoramic unfolded images of the borehole. In one embodiment of this invention, a semantic segmentation model pre-trained on a borehole image dataset is used to perform pixel-level classification of the images. Preferably, the semantic segmentation model adopts the U-Net architecture, with an input image resolution of 512×512 pixels, and output classification categories including five types: intact rock walls, open fractures, closed fractures, seepage marks, and lithological boundaries. A sliding window of 0.5m is used along the borehole depth direction, and the percentage of fracture-type pixels within the window is calculated as the fracture rate (CR), and the percentage of seepage-type pixels is calculated as the seepage rate (SR). The attitude and width of the identified major fractures are extracted. The attitude is obtained through image geometric correction and 3D reconstruction, and the width is obtained through the conversion relationship between pixel width and actual size. Finally, a visual feature profile along the borehole depth is generated, including information such as borehole number, starting mileage, mileage conversion values ​​for each depth point, fracture rate, seepage rate, and dominant fracture dip angle.

[0028] Through the collection and structured processing of the four types of data described above, this step constructs a multi-source geological database with tunnel mileage as a unified index, providing a standardized data foundation for subsequent risk classification and fusion identification. In one embodiment of this invention, the database adopts a relational storage architecture, with the mileage field as the primary index and a uniform precision of 0.1m. Data from different sources are aligned in time and space through the mileage index, facilitating subsequent cross-source data querying and fusion analysis.

[0029] Step S2: Dynamic Risk Classification and Verification Strategy Matching. The core innovation of this step lies in constructing a risk classification-driven intelligent matching mechanism for verification strategies, changing the traditional fixed-combination forecasting model and achieving optimized allocation of forecasting resources. Based on the long-range geophysical forecasting results obtained in Step S1, a quantitative risk assessment model is constructed to calculate the initial risk index, and differentiated short-range verification strategies are automatically triggered according to the risk level.

[0030] Initial risk index The calculation employs a multi-factor weighted model. In one embodiment of the present invention, the formula for calculating the initial risk index is: , in, The initial risk index is dimensionless and ranges from 0 to 100, with higher values ​​indicating greater risk. This is a normalized value for the intensity of geophysical anomalies. It is dimensionless and ranges from 0 to 100. It is calculated based on the deviation of the peak value of the reflection intensity curve from the background value. When the deviation exceeds 3 times the standard deviation, the value is 100. The normalized value of the abnormal section length is dimensionless and ranges from 0 to 100. It is calculated based on the ratio of the abnormal section length to the preset reference length (set to 50m in this embodiment). When the ratio exceeds 1, the value is 100. The deviation from the designed geological conditions is dimensionless and ranges from 0 to 100. It is obtained by expert system evaluation based on the degree of difference between the actual detection results and the expected geological conditions in the design documents. , and These are the weighting coefficients for the three factors, which in this embodiment are 0.5, 0.3, and 0.2 respectively, and the sum of the three weighting coefficients equals 1.

[0031] The risk level is divided into four levels. In one embodiment of the present invention, three risk level thresholds are set: the first threshold... Set to 30, the second threshold Set to 50, the third threshold Set it to 75. When the initial risk index... Less than the first threshold When the initial risk index is greater than or equal to the first threshold and less than the second threshold, it is classified as R0 (low risk); when the initial risk index is greater than or equal to the second threshold and less than the third threshold, it is classified as R1 (low to medium risk); when the initial risk index is greater than or equal to the second threshold and less than the third threshold, it is classified as R2 (medium risk); and when the initial risk index is greater than or equal to the third threshold, it is classified as R3 (high risk). Preferably, the above thresholds can be dynamically adjusted based on engineering experience and historical data to adapt to the characteristics of different geological environments.

[0032] The verification strategy matching engine automatically triggers corresponding differentiated short-range verification strategies based on the risk level. In one embodiment of the present invention, the verification strategy is configured as follows: For low-risk sections of R0 and R1 levels, a routine verification strategy is triggered. This strategy includes two components: real-time analysis of drilling parameters and face point cloud scanning. Real-time analysis of drilling parameters involves continuously monitoring the trends of the drilling specific energy index and fluctuation coefficient during subsequent tunneling, providing timely warnings when abnormal fluctuations occur. Face point cloud scanning refers to a routine 3D laser scan performed after each excavation cycle to update the structural surface feature data.

[0033] For R2-level medium-risk sections, an enhanced verification strategy is triggered. This enhanced verification strategy adds two components to the conventional strategy: first, short-range ground-penetrating radar scanning, using high-frequency ground-penetrating radar to detect areas 20 to 40 meters ahead of the anomaly section, with a center frequency of 400 MHz and a vertical resolution better than 0.1 m; second, strengthened analysis of drilling parameters, tightening the monitoring thresholds for drilling parameters by 20%, i.e., triggering an early warning when the drilling specific energy index decreases by 15% or the fluctuation coefficient exceeds 0.24.

[0034] For R3-level high-risk sections, the highest-level verification strategy is triggered. This strategy, based on the enhanced verification strategy, immediately implements advanced horizontal drilling and borehole wall imaging for deterministic verification. Preferably, the advanced horizontal drilling depth is no less than 30m and the borehole diameter is 76mm, with drilling parameters recorded simultaneously during the drilling process. After drilling is completed, a borehole camera system acquires unfolded images of the entire borehole section, which are automatically interpreted by the visual analysis module in step S1.

[0035] Taking this embodiment as an example, assuming that the seismic wave method predicts a strong reflection anomaly in the section from DK158+220 to DK158+260, with a normalized reflection intensity value of 85, and a normalized value of 80 corresponding to an anomaly section length of 40m, and an assessment value of 60 for the degree of deviation from the design geological conditions, the initial risk index is calculated by substituting these values ​​into the above formula. The result is 0.5 multiplied by 85 plus 0.3 multiplied by 80 plus 0.2 multiplied by 60, which equals 78.5. Since 78.5 is greater than or equal to the third threshold of 75, the system marks this section as R3 level high risk. The decision engine automatically generates the instruction: "When the tunnel face reaches DK158+210, immediately initiate advanced horizontal drilling to conduct deterministic verification of the section from DK158+210 to DK158+240, and at the same time strengthen the real-time monitoring and analysis of the drilling parameters of all system boreholes after DK158+210."

[0036] Through the dynamic risk grading and verification strategy matching mechanism described above, this invention achieves optimized allocation of forecast resources, fully verifies high-risk areas, avoids over-exploration of low-risk areas, and optimizes economy while improving safety.

[0037] Step S3: Multi-level, multi-source information fusion and intelligent identification. This step is the core technology of this invention. It employs a layered fusion strategy to organically combine information of different scales, precisions, and types to form a comprehensive identification result. The key innovation of this invention lies in proposing an adaptive confidence-weighted fusion algorithm, which can dynamically adjust the fusion weights according to the degree of evidence conflict between various data sources, effectively solving the problem of traditional evidence theory failing under high conflict conditions.

[0038] First, feature-level fusion is performed, aligning and fusing long-range forecast features with short-range verification features and face exposure features under a unified mileage coordinate system. In one embodiment of the invention, a unified spatial grid is established with mileage intervals of 0.1m. Data from different sources and with different sampling intervals are resampled using linear interpolation or nearest neighbor interpolation methods to align all data to the same mileage point. Subsequently, each feature is normalized, and a minimum-maximum normalization method is used to map all feature values ​​to the interval between 0 and 1. Finally, a multi-dimensional feature matrix X is constructed, where rows correspond to different mileage points and columns correspond to different feature dimensions, including seismic wave reflection intensity, ground-penetrating radar reflection intensity, drilling energy index, joint density, and fracture rate.

[0039] Building upon feature-level fusion, this invention employs an adaptive confidence-weighted fusion algorithm for decision-level fusion. The core idea of ​​this algorithm is to dynamically adjust the fusion weights based on the degree of evidence conflict between data sources. Data sources with high conflict are weighted less, while those with low conflict are weighted more, thereby improving the reliability of the fusion results.

[0040] Conflict of Evidence The calculation formula is: , in, , represents the degree of evidence conflict between data source i and data source j, is dimensionless, and ranges from 0 to 1. The closer the value is to 1, the more contradictory the identification results of the two data sources are; K is the total number of geological categories. In this embodiment, K equals 6, including six categories: intact rock mass, slightly fractured, moderately fractured, severely fractured, water-rich zone, and fault. Let be the identification probability of data source i for the k-th geological condition. It is dimensionless and ranges from 0 to 1. It is obtained from the output of the independent preliminary identification model of each data source. Let be the probability of data source j identifying the k-th type of geological condition; The smaller of the two probabilities is taken, and the summation term represents the degree of overlap between the identification results of the two data sources. The higher the degree of overlap, the lower the degree of conflict.

[0041] The fusion weight of each data source is calculated based on the degree of evidence conflict. The calculation formula is: , in, The fusion weight of the i-th data source is dimensionless and ranges from 0 to 1. The sum of the weights of all data sources is equal to 1. N is the total number of data sources. In this embodiment, N is equal to 4, including four types: geophysical data, drilling parameters, point cloud data, and borehole images. The average conflict degree between the i-th data source and all other data sources is calculated by taking the arithmetic mean of the conflict degrees between the i-th data source and all other data sources. This is a conflict attenuation factor, dimensionless, with a value ranging from 0.5 to 2.0, used to control the influence of conflict degree on weight. In this embodiment, the value is 1.2. The formula is a natural exponential function; the denominator is a normalization factor to ensure that the sum of all weights equals 1. The technical effect of this formula is that when the average conflict between a data source and other data sources is high, its exponential term is smaller, resulting in a lower fusion weight; conversely, it receives a higher weight.

[0042] In one embodiment of the present invention, when the average conflict degree among all data sources exceeds a preset conflict threshold (set to 0.6 in this embodiment), the system automatically triggers a manual review process, prompting engineering technicians to intervene and make a judgment.

[0043] The weighted multidimensional feature matrix is ​​input into the fusion identification model for intelligent identification. In one embodiment of the present invention, the fusion identification model adopts a multi-branch deep neural network structure, including three parallel branches: The first branch is a sequence feature processing branch, which uses a one-dimensional convolutional neural network to process sequence features that change along the mileage, including seismic wave reflection intensity sequence, ground-penetrating radar reflection intensity sequence, and drilling specific energy index sequence. Preferably, the kernel size of the one-dimensional convolutional layer is 3, the number of convolutional kernels is 64, and the activation function is a modified linear unit. Local patterns and changing trends are captured through multi-layer convolution and pooling operations. The second branch is an image feature processing branch, which uses an image convolutional neural network to process the current face image and typical borehole images to extract texture, color, and edge features. Preferably, a pre-trained residual network is used as a feature extractor to output a 256-dimensional feature vector. The third branch is a context feature processing branch, which uses a fully connected network to process discrete context features (such as burial depth and design lithology coding) and statistical features (such as the mean and variance of each feature). The outputs of the three branches are concatenated and then fully connected. Finally, the Softmax layer outputs the probability of occurrence of type K geological hazards. .

[0044] In one embodiment of the invention, a dynamic Bayesian network is also employed for auxiliary inference to enhance the model's interpretability and robustness to incomplete data. The dynamic Bayesian network consists of two layers: evidence nodes and hypothesis nodes. The evidence nodes include four observed variables: strong geophysical reflection, sudden drop in mechanical parameters, fracture development, and signs of water seepage. The hypothesis nodes include four target variables: intact rock mass, fractured zone, water-rich body, and fault. The network's conditional probability table is obtained from historical case data through maximum likelihood estimation and can be continuously updated based on new data. The currently observed evidence is input into the dynamic Bayesian network for probabilistic inference to obtain the Bayesian inference probability.

[0045] Final comprehensive risk index It is obtained through an adaptive weighted fusion of deep learning probabilities and Bayesian inference probabilities, and the calculation formula is as follows: , in, This is a comprehensive risk index, dimensionless, with a value range of 0 to 100. The higher the value, the greater the risk of geological disasters. The risk index output by the deep learning model is obtained by weighted summation of the probabilities of various geological disasters, with the weights set according to the severity of each disaster. The risk index output by the Bayesian network inference is calculated using the same method as the risk index for deep learning. The fusion coefficient is dimensionless and ranges from 0 to 1. In this embodiment, it is set to 0.6, meaning the deep learning result accounts for 60% of the weight and the Bayesian network result accounts for 40%. Preferably, the fusion coefficient can be dynamically adjusted based on the prediction accuracy of the two models on historical data, with the model with higher accuracy receiving a larger weight.

[0046] Taking this embodiment as an example, the segment from DK158+215 to DK158+225 was fused and identified. Geophysical data showed that the normalized value of seismic wave reflection intensity in this segment was 0.85, and the normalized value of ground-penetrating radar reflection intensity was 0.90. Drilling parameters showed that the drilling specific energy index plummeted from 300 MJ / m³ to 100 MJ / m³, with a normalized value of 0.15, and the fluctuation coefficient surged to 0.65. Point cloud data showed that the joint density Jv was 4.5 joints / m³, with a normalized value of 0.70. The deep learning model output a probability of 0.78 for a water-rich fractured zone, corresponding to a risk index of... The value is 78. The posterior probability of the water-rich, fragmented zone output by the Bayesian network inference is 0.82, corresponding to a risk index of 78. The value is 82. Substituting this into the fusion formula yields the comprehensive risk index. The result is 0.6 multiplied by 78 plus 0.4 multiplied by 82, which equals 79.6. Since 79.6 is greater than the threshold of 75 and the evidence from multiple sources is highly consistent, the system raises the risk level of this section from R2 to R3 and marks the disaster type as a high-probability water-rich fractured zone.

[0047] Step S4: Prediction Result Output and Feedback Optimization. This step realizes the visualization output of prediction results, the issuance of early warnings, and the continuous optimization of the model based on excavation and exposure, and builds a closed-loop feedback mechanism of prediction-exposure-comparison-optimization.

[0048] Risk levels are classified and early warnings are issued based on the comprehensive risk index output in step S3. In one embodiment of the present invention, the risk level classification standard is consistent with the threshold setting in step S2, that is, a comprehensive risk index of less than 30 is low risk, 30 to 50 is low-to-medium risk, 50 to 75 is medium risk, and greater than or equal to 75 is high risk. Early warning information is simultaneously released through multiple channels, including highlighting risk areas with different colors on the 3D visualization interface of the building information model, pushing early warning messages to the mobile terminals of relevant personnel, and automatically generating disposal suggestion reports. Preferably, the disposal suggestion report includes the risk level, disaster type, suggested measures (such as adjusting support parameters, starting pre-grouting, preparing drainage equipment, etc.), and subsequent monitoring requirements.

[0049] The calculation of prediction error is the foundation of feedback optimization. In one embodiment of this invention, the actual geological conditions revealed during excavation are used as the true value label and compared with the prediction results. The actual geological conditions are recorded using standardized geological logging forms, filled out by the on-site geological engineer after each excavation cycle. The records include the exposed mileage, surrounding rock grade, degree of joint development, seepage status, and actual hazard type. Prediction error. The calculation formula is: , in, Let be the prediction error for the t-th excavation cycle, dimensionless, and ranging from 0 to 100; Let be the predicted risk index for the t-th cycle; The actual risk index for the t-th cycle is calculated according to standardized scoring rules based on the revealed geological conditions. Preferably, the classification error, i.e., whether the predicted hazard type matches the actual revealed type, can also be calculated.

[0050] Online incremental learning is a key mechanism for achieving continuous model evolution. In one embodiment of this invention, the implementation process of online incremental learning is as follows: First, the multi-source data, model prediction results, and actual exposure conditions of each excavation cycle are packaged into a training sample and stored in a historical database. The sample format includes input features (multi-dimensional feature matrix), model output (various probabilities and comprehensive risk indices), and ground truth labels (actual geological type and actual risk index). Second, new training samples are extracted from the historical database according to a preset cycle (once a week in this embodiment), typically ranging from 50 to 200. Then, an incremental learning algorithm is used to fine-tune the parameters of the fusion identification model. Preferably, incremental learning employs an elastic weight consolidation method, optimizing the loss of new samples while maintaining the memory of old knowledge, avoiding catastrophic forgetting. Learning rate. The value was set to 0.01, and the number of iterations was set to 10. The fine-tuning process focused on optimizing categories with low accuracy under the current geological conditions, improving the recognition ability of these categories by adjusting the classification layer weights.

[0051] Simultaneously, the conditional probabilities of each evidence node in the dynamic Bayesian network are fine-tuned based on the statistical results of the forecast accuracy. In one embodiment of the invention, the conditional probability table is updated using a Bayesian posterior estimation method. Specifically, the joint frequency of each evidence variable and the target variable in the newly added samples is statistically analyzed and then weighted and averaged with the original conditional probability table, with the weights determined by the ratio of the number of new and old samples. This update mechanism makes the inference results of the Bayesian network more consistent with the actual geological characteristics of the current work area.

[0052] Through the aforementioned closed-loop feedback mechanism, this invention enables the forecasting model to continuously adapt and optimize as the project progresses. With the advancement of excavation mileage and the accumulation of training samples, the model's adaptability to the current tunnel geological environment continuously improves, and the forecast accuracy gradually increases. In one embodiment of this invention, application in a high-speed railway tunnel project shows that after incremental learning over approximately 500m of mileage, the overall forecast accuracy increased from the initial 75% to 89%, the false alarm rate decreased from 18% to 7%, and the missed alarm rate decreased from 12% to 4%, verifying the effectiveness of the closed-loop feedback optimization mechanism.

[0053] This invention also provides a multi-source information fusion intelligent prediction system for tunnel geological hazards that implements the above-mentioned method. In one embodiment of this invention, the system adopts a layered distributed architecture, including four core functional modules: a data acquisition module, a risk classification module, a fusion identification module, and a feedback optimization module, as well as four deployment layers: a perception and acquisition layer, an edge computing and network layer, a platform and algorithm layer, and an application and interaction layer.

[0054] The data acquisition module is deployed at the tunnel construction site and is responsible for systematically collecting multi-source heterogeneous data generated during tunnel construction. This data is then uniformly converted into a structured data list with tunnel mileage as the key index, constructing a spatiotemporally aligned multi-source geological database. The data acquisition module includes four sub-units: a geophysical data interface unit with a protocol converter, connecting to the data output ports of devices such as seismic wave detectors, ground-penetrating radar, and transient electromagnetic instruments, enabling automatic capture and uploading of raw data files and report files; a drilling parameter acquisition unit with an industrial-grade IoT gateway installed on the programmable logic controller of the drilling rig, which collects propulsion pressure, rotation pressure, impact pressure, and drill pipe displacement signals in real time at a frequency of 10Hz via a standard communication protocol, and binds the time-series data to the absolute mileage of the borehole through a coordinate transformation model; and a 3D sensing unit with an explosion-proof 3D laser scanner fixedly installed on the sidewall approximately 50m from the working face, automatically scanning after each cycle of blasting and muck removal, and transmitting point cloud data back via a wireless network; simultaneously, an explosion-proof high-definition intelligent camera is installed at this location to continuously monitor the working face status. The borehole imaging unit is equipped with a digital borehole camera system for acquiring panoramic images of the borehole during advanced horizontal drilling.

[0055] The edge computing and network layer is deployed in the control room at the tunnel entrance, including high-performance edge servers and industrial switches. All perception layer data is aggregated to the edge server via a hybrid wired and wireless network for real-time preprocessing, including data format conversion, outlier removal, and timestamp synchronization, to reduce cloud transmission pressure and meet real-time requirements. The edge server also handles part of the workload for computationally intensive tasks in step S1, such as intelligent parsing of geophysical reports, calculation of drilling parameter features, and identification of point cloud structural surfaces.

[0056] The risk grading module is connected to the data acquisition module and is used to construct a quantitative risk assessment model based on multi-source data, calculate the initial risk index of the section to be predicted, determine the risk level, and automatically trigger the corresponding differentiated short-range verification strategy through the verification strategy matching engine. In one embodiment of the present invention, the risk grading module adopts a hybrid architecture combining a rule engine and a machine learning model. The rule engine is responsible for determining the risk level and triggering the verification strategy according to preset threshold judgment logic. The machine learning model is responsible for learning a more refined risk quantification function based on historical data. The outputs of the two are combined through a voting mechanism to improve the robustness of the judgment. The verification strategy matching engine maintains a strategy library, stores the verification strategy templates corresponding to each risk level, and instantiates and generates specific execution instructions based on current working condition parameters (such as the working face position, available equipment status, etc.).

[0057] The fusion identification module is connected to the data acquisition module and the risk classification module. It performs spatiotemporal alignment and normalization processing on long-range forecast features and short-range verification features to construct a multi-dimensional feature matrix. An adaptive confidence-weighted fusion algorithm and fusion identification model are used to output the probability of occurrence and comprehensive risk index of various geological disasters. In one embodiment of the invention, the core algorithm of the fusion identification module is deployed on a cloud server or enterprise private cloud server, receiving structured data from edge servers and running the multi-branch deep neural network and dynamic Bayesian network model described in step S3. Model inference adopts a combination of batch processing and real-time processing; batch processing is used for routine forecasts to improve throughput, while real-time processing is used for emergency warnings to ensure response speed. The fusion identification module also includes a model management submodule, responsible for model version control, model deployment, and model performance monitoring.

[0058] The feedback optimization module is connected to the fusion identification module and is used to issue early warnings based on a comprehensive risk index. It compares the actual geological conditions revealed by excavation with the forecast results to calculate the prediction error, and performs online incremental learning and parameter adaptive optimization of the fusion identification model based on the prediction error. In one embodiment of the invention, the feedback optimization module includes three components: an early warning issuance submodule, a data retrieval submodule, and an incremental learning submodule. The early warning issuance submodule simultaneously releases early warning information through multiple channels, including web browser interfaces, mobile applications, and SMS notifications. The data retrieval submodule provides a standardized geological logging interface for field engineers to input actual exposure conditions and automatically matches and correlates the input data with the forecast results. The incremental learning submodule automatically executes the model fine-tuning process described in step S4 according to a preset cycle and generates a training report for technical personnel to review.

[0059] The application and interaction layer provides on-site engineers, project managers, and back-end experts with a visual interface, risk warnings, and decision-making suggestions through web browsers and mobile applications. In one embodiment of the invention, the visual interface is built based on Building Information Modeling (BIM) technology, overlaying the 3D tunnel model with the forecast results, using different colors to identify sections with different risk levels, and supporting interactive operations such as mileage query, historical data review, and trend analysis. The mobile application supports mainstream mobile operating systems and provides functions such as early warning message push, on-site situation feedback, and voice annotation, facilitating on-site personnel to obtain information and report data anytime, anywhere.

[0060] Through the aforementioned system architecture, this invention achieves a complete closed loop from data perception, intelligent processing, fusion analysis to decision feedback, providing an engineering-based and systematic solution for geological disaster prediction during tunnel construction. The system's modular design facilitates tailoring and expansion according to specific engineering needs, while the edge-cloud collaborative deployment architecture balances real-time performance and computing power. The closed-loop feedback mechanism ensures continuous system optimization.

[0061] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A multi-source information fusion method for intelligent prediction of tunnel geological hazards, characterized in that, Includes the following steps: S1. Multi-source heterogeneous data acquisition and structured conversion steps: systematically collect multi-source heterogeneous data generated during tunnel construction, and convert it into a structured data list with tunnel mileage as the key index to build a spatiotemporally aligned multi-source geological database. S2. Dynamic risk grading and verification strategy matching steps: Based on the multi-source data obtained in step S1, construct a quantitative risk assessment model and calculate the initial risk index of the segment to be predicted. The risk level is determined based on the comparison between the initial risk index and the preset risk level threshold. Based on the risk level, the verification strategy matching engine automatically triggers the corresponding differentiated short-range verification strategy. Low risk level triggers the regular verification strategy, medium risk level triggers the enhanced verification strategy, and high risk level triggers the highest level verification strategy. S3, Multi-level and multi-source information fusion and intelligent identification steps: The long-range forecast features and short-range verification features obtained in steps S1 and S2 are spatiotemporally aligned and normalized in a unified mileage coordinate system to construct a multi-dimensional feature matrix. An adaptive confidence-weighted fusion algorithm is used to calculate the fusion weight of each data source, where the fusion weight is dynamically adjusted according to the degree of evidence conflict between the data sources; The multidimensional feature matrix is ​​input into the fusion identification model, which outputs the probability of occurrence and comprehensive risk index of various geological disasters. S4. Prediction Result Output and Feedback Optimization Steps: Based on the comprehensive risk index output in step S3, risk level classification and early warning issuance are carried out. The actual geological conditions revealed by the excavation are used as the true value label and compared with the prediction results to calculate the prediction error; Based on the prediction error, the fusion identification model is subjected to online incremental learning and adaptive parameter optimization to achieve continuous evolution of forecasting capabilities.

2. The intelligent prediction method for tunnel geological hazards based on multi-source information fusion according to claim 1, characterized in that, In step S1, the multi-source heterogeneous data includes geophysical data, drilling parameter data, three-dimensional point cloud data of the working face, and borehole visual data; the geophysical report is intelligently analyzed to extract abnormal mileage intervals and physical property parameters, and a list of geophysical features is generated. The drilling parameters are filtered and feature-calculated to extract the drilling specific energy index and generate a list of mechanical responses. The process of identifying structural surfaces and calculating attitudes of point clouds at the working face to generate a list of structural surface features and intelligently parsing geophysical reports includes: using optical character recognition technology to convert geophysical reports into a processable digital format; The reflection profile image is binarized and the coordinate axis is identified. The reflection intensity curve is extracted through the pixel-mileage mapping relationship. A pre-trained named entity recognition model is used to extract mileage intervals, anomaly features and inferred conclusions from the text conclusions to generate a structured geophysical feature list.

3. The intelligent prediction method for tunnel geological hazards based on multi-source information fusion according to claim 1, characterized in that, In step S1, the drilling specific energy index is calculated as follows: taking each preset drilling length as an analysis unit, the average thrust, average torque, average drilling speed and drill rod rotation speed within the analysis unit are calculated; the drilling specific energy index is calculated based on the average thrust, average torque, average drilling speed, drill rod rotation speed and borehole cross-sectional area; at the same time, the thrust fluctuation coefficient is calculated as an evaluation index of rock mass uniformity.

4. The intelligent prediction method for tunnel geological hazards based on multi-source information fusion according to claim 1, characterized in that, In step S2, the initial risk index is obtained by weighting the intensity of geophysical anomalies, the length of anomaly sections, and the degree of deviation from the designed geological conditions. The risk level thresholds include a first threshold, a second threshold, and a third threshold. When the initial risk index is less than the first threshold, it is a low risk level; when the initial risk index is greater than or equal to the first threshold and less than the second threshold, it is a medium-low risk level; when the initial risk index is greater than or equal to the second threshold and less than the third threshold, it is a medium risk level; and when the initial risk index is greater than or equal to the third threshold, it is a high risk level.

5. The intelligent prediction method for tunnel geological hazards based on multi-source information fusion according to claim 1, characterized in that, In step S3, the adaptive confidence-weighted fusion algorithm includes: calculating the degree of evidence conflict between any two data sources, the degree of evidence conflict being determined based on the degree of difference between the two data sources in their identification results for the same mileage segment; calculating the fusion weight of each data source based on the degree of evidence conflict, wherein data sources with higher conflict with other data sources are assigned lower fusion weights; and triggering a manual review process when the average degree of conflict between all data sources exceeds a preset conflict threshold.

6. The intelligent prediction method for tunnel geological hazards based on multi-source information fusion according to claim 1, characterized in that, In step S3, the fusion identification model adopts a multi-branch deep neural network structure, including: the first branch uses a one-dimensional convolutional neural network to process the sequence features that change along the mileage, capturing local patterns and trends; the second branch uses an image convolutional neural network to process the face image and borehole image, extracting texture and color features; the third branch uses a fully connected network to process discrete context features and statistical features; the outputs of the three branches are concatenated and then output the probability of occurrence of various geological disasters through a fully connected layer and a classification layer.

7. The intelligent prediction method for tunnel geological hazards based on multi-source information fusion according to claim 1, characterized in that, Step S3 also includes using a dynamic Bayesian network for auxiliary reasoning. The evidence nodes of the dynamic Bayesian network include strong geophysical reflection, sudden drop in mechanical parameters, fracture development, and signs of water seepage. The assumed nodes include intact rock mass, fracture zone, water-rich body, and fault. The currently observed evidence is input into the dynamic Bayesian network for probabilistic reasoning to obtain the Bayesian reasoning probability. The probability output by the deep neural network is adaptively weighted and fused with the Bayesian reasoning probability to obtain the final comprehensive risk index.

8. The intelligent prediction method for tunnel geological hazards based on multi-source information fusion according to claim 1, characterized in that, In step S4, online incremental learning includes: packaging multi-source data, model prediction results, and actual exposure conditions of each excavation cycle into training samples and storing them in a historical database; extracting new training samples from the historical database according to a preset cycle; using an incremental learning algorithm to fine-tune the parameters of the fusion identification model, focusing on optimizing the identification weights under the current geological conditions; and fine-tuning the conditional probabilities of each evidence node in the dynamic Bayesian network based on the statistical results of the prediction accuracy.

9. The intelligent prediction method for tunnel geological hazards based on multi-source information fusion according to claim 1, characterized in that, In step S2, the conventional verification strategy includes real-time analysis of drilling parameters and point cloud scanning of the drilling face; the enhanced verification strategy includes adding short-range ground-penetrating radar scanning and enhanced analysis of drilling parameters on the basis of the conventional verification strategy; the highest level verification strategy includes immediately implementing advanced horizontal drilling and borehole wall imaging on the basis of the enhanced verification strategy to conduct deterministic verification.

10. A multi-source information fusion tunnel geological hazard intelligent prediction system, used to implement the multi-source information fusion tunnel geological hazard intelligent prediction method according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to systematically collect multi-source heterogeneous data generated during tunnel construction and convert it into a structured data list with tunnel mileage as the key index, thus constructing a spatiotemporally aligned multi-source geological database. The data acquisition module includes a geophysical data interface unit, a drilling parameter acquisition unit, a 3D sensing unit, and a borehole imaging unit. The risk grading module, connected to the data acquisition module, is used to build a quantitative risk assessment model based on multi-source data, calculate the initial risk index of the segment to be predicted, determine the risk level, and automatically trigger the corresponding differentiated short-distance verification strategy through the verification strategy matching engine. The fusion identification module, connected to the data acquisition module and the risk classification module, is used to perform spatiotemporal alignment and normalization of long-range forecast features and short-range verification features to construct a multi-dimensional feature matrix. It uses an adaptive confidence-weighted fusion algorithm and fusion identification model to output the occurrence probability and comprehensive risk index of various geological disasters. The feedback optimization module, connected to the fusion identification module, is used to issue early warnings based on the comprehensive risk index. It compares the actual geological conditions revealed by the excavation with the forecast results to calculate the prediction error, and performs online incremental learning and parameter adaptive optimization of the fusion identification model based on the prediction error.

Images