Multi-source information driven tunnel surrounding rock grade identification method and system
Patent Information
- Application Number
- CN202610808027.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-05
- Publication Date
- 2026-09-25
AI Technical Summary
[0003]然而，当前隧道工程施工过程中，围岩等级判定工作多依赖现场技术人员人工完成，判定结果受个人从业经验、专业能力的主观影响显著，不同人员的判定结果一致性不足，且人工判定的效率难以匹配隧道工程快速施工的节奏要求，现有围岩等级智能化判定方案，多采用单一类型或少量数据源开展分析，无法全面覆盖隧道围岩的宏观地质背景与施工过程中的动态变化特征，判定结果的稳定性难以保障，同时，现有智能化方案未构建完整的判定、校验、优化闭环，无法结合现场实测结果持续优化判定能力，面对不同地质条件的隧道工程，判定性能出现显著波动，难以适配复杂地质场景的工程应用需求
一、本发明通过多源隧道地质数据的协同应用，将隧道前期设计阶段的宏观地质背景数据，与施工过程中实时采集的超前地质预报、现场钻孔、隧道断面图像数据相结合，构建与不同数据类型相匹配的特征提取路径与跨源融合机制，完整覆盖隧道围岩从宏观地质背景到局部实时状态的全维度信息，消除单一数据源带来的信息局限性，为围岩等级的精准判定提供全面的信息支撑，整体方案贴合隧道工程从勘察设计到现场施工的全流程作业逻辑，具备较高的工程落地适配性。
Smart Images
Figure CN122818104A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of tunnel engineering technology, specifically to a method and system for identifying the grade of surrounding rock in tunnels driven by multi-source information. Background Technology
[0002] In the infrastructure construction system of highways, railways, water conservancy and other infrastructure, tunnel engineering is the core structure for traversing complex terrains such as mountains and hills. It is also a key engineering form to ensure smooth traffic routes, shorten travel distances and reduce ecological disturbance. As infrastructure construction continues to extend to the mountainous areas of central and western China, the scale and number of long-distance, deep-buried tunnel projects that traverse complex geological units are constantly increasing. The geological conditions faced by tunnel construction are becoming increasingly complex. The determination of the surrounding rock grade of the tunnel runs through the entire tunnel construction cycle. It is the core basis for tunnel support scheme design, construction procedure planning and on-site safety management. It directly determines the construction safety, cost control and the stability of the tunnel project throughout its entire life cycle. The industry regards the accurate and timely determination of the surrounding rock grade during tunnel construction as the core management link for tunnel engineering safety management and high-quality construction.
[0003] However, in current tunnel construction, the determination of surrounding rock grade largely relies on manual work by on-site technicians. The determination results are significantly influenced by the subjective factors of individual experience and professional ability, resulting in insufficient consistency among different personnel. Furthermore, the efficiency of manual determination is difficult to match the fast-paced construction requirements of tunnel projects. Existing intelligent surrounding rock grade determination schemes often use a single type or a small number of data sources for analysis, which cannot fully cover the macroscopic geological background of the tunnel surrounding rock and the dynamic changes during the construction process. The stability of the determination results is difficult to guarantee. At the same time, existing intelligent schemes have not constructed a complete closed loop of determination, verification, and optimization, and cannot continuously optimize the determination capability based on on-site measurement results. When facing tunnel projects with different geological conditions, the determination performance fluctuates significantly, making it difficult to adapt to the engineering application needs of complex geological scenarios. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for identifying the surrounding rock grade of tunnels driven by multi-source information. This invention combines macroscopic geological background data from the early design stage of tunnels with real-time geological forecasts, on-site borehole data, and tunnel cross-section image data collected during construction. It constructs feature extraction paths and cross-source fusion mechanisms that match different data types, fully covering all dimensions of information on the surrounding rock of tunnels from macroscopic geological background to local real-time status. This eliminates the information limitations caused by a single data source and provides comprehensive information support for the accurate determination of the surrounding rock grade. The overall solution fits the entire process of tunnel engineering from survey and design to on-site construction and has high adaptability to engineering implementation.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a multi-source information-driven method for identifying the grade of surrounding rock in tunnels, the specific steps of which are as follows: Data acquisition and preprocessing: Acquire design surrounding rock grade data, TSP advanced geological prediction data, high-definition image data of tunnel cross section, and on-site borehole parameter data of the target tunnel construction section, and preprocess them respectively to obtain standardized multi-source datasets; Feature extraction and cross-source fusion: Features are extracted from each data point in the standardized multi-source dataset to obtain the corresponding feature vectors. Cross-source feature fusion is then performed on all feature vectors to obtain a global fused feature vector. Rock grade prediction output: Input the global fused feature vector into the pre-trained deep learning model for rock grade recognition, and output the predicted rock grade of the target tunnel construction section; Prediction result verification: Obtain the true value of the surrounding rock grade determined by on-site measurement of the target tunnel construction section, compare the predicted surrounding rock grade result with the true value of the surrounding rock grade, calculate and output the prediction accuracy value based on the comparison result; Closed-loop optimization: The standardized multi-source dataset, the true value of the surrounding rock grade, and the prediction accuracy value are combined to form an incremental training sample set. Based on the preset iteration triggering conditions, the incremental training sample set is input into the deep learning model for surrounding rock grade recognition, and incremental self-learning training is performed to update the network weights of the deep learning model for surrounding rock grade recognition, thus completing the iterative optimization.
[0006] Furthermore, in the data acquisition and preprocessing, design surrounding rock grade data is obtained from the construction drawing design documents issued during the early-stage tunnel engineering survey; TSP (Transient Seismic Wave Reflection) advanced geological prediction data is obtained from the advanced geological prediction equipment during tunnel construction; high-definition image data of the tunnel cross-section is acquired on-site using a high-definition industrial camera; and on-site drilling parameter data is acquired in real time using advanced geological borehole measurement-while-drilling equipment. The design surrounding rock grade data includes the surrounding rock grade classification results given in the construction drawing design documents, the corresponding geological lithology, the design value of the rock mass integrity coefficient, and the design value of the uniaxial compressive strength of the surrounding rock; the TSP advanced geological prediction data includes the P-wave velocity of the surrounding rock, the spatial location of the reflection interface, and rock mass integrity parameters. The data includes the distribution range of adverse geological bodies and the elastic modulus of the surrounding rock; high-definition image data of the tunnel cross-section, including full-range RGB images of the tunnel face and cross-sectional images of the surrounding rock before the initial support around the tunnel; and field drilling parameter data, including drilling speed, drill rod torque, drilling pressure, flushing fluid flow rate, core recovery rate, and rock fragmentation degree records of advanced geological boreholes. During preprocessing, outliers were removed from numerical data using the 3σ criterion and then normalized. Wavelet transform was used to denoise the time-series data, followed by time alignment and normalization. Dark channel prior dehazing was used on the image data, followed by size unification and pixel value normalization. All data were mapped to the numerical range of 0-1 to obtain a standardized multi-source dataset.
[0007] Furthermore, in the feature extraction and cross-source fusion, different feature extraction paths are adopted for different data in the standardized multi-source dataset. For static numerical data, a three-layer fully connected neural network is used for feature extraction. Each fully connected layer is followed by a ReLU activation function to output a fixed-dimensional numerical feature vector. The static numerical data is the design surrounding rock grade data. For time-series dynamic data, a three-layer causal convolutional network is used to extract local time-series features. The extracted local features are then connected to a two-layer long short-term memory network to extract long-distance time-series dependent features, outputting a fixed-dimensional time-series feature vector. The time-series dynamic data is TSP advanced geological prediction data and on-site borehole parameter data. For high-definition image data of tunnel cross sections, a ResNet18 backbone convolutional neural network is used. After removing the classification layer at the end, image features are extracted, outputting a fixed-dimensional image feature vector.
[0008] Furthermore, in the feature extraction and cross-source fusion, the cross-source feature fusion process constructs a global fusion feature vector through a cross-attention weighted fusion formula for surrounding rock features. The cross-attention weighted fusion formula for surrounding rock features is as follows: ,in, This represents the global fusion feature vector, where i is the index of the feature vector, taking values of 1, 2, and 3, corresponding to numerical, time-series, and image data, respectively. Let i be the i-th eigenvector. For the fusion weight coefficients corresponding to the i-th feature vector, all The summation result is 1, which is obtained through calculation using the formula: ,in, This serves as the engineering benchmark weighting coefficient, set according to the contribution ratio of various geological exploration methods in the application scenario of tunnel surrounding rock classification engineering. denoted as data validity and completeness, and denoted as the ratio of valid data volume after outlier removal during preprocessing to the total collected data volume. j is the summation variable, corresponding to i.
[0009] Furthermore, in the grade prediction output, the deep learning model for surrounding rock grade identification adopts a multi-branch input single-output classification network architecture. The input layer has four parallel input branches, corresponding to the input interfaces of four types of data. The intermediate layer has a feature extraction sub-network and a cross-attention fusion sub-network that match the data. The output layer has six output nodes, corresponding to tunnel surrounding rock grades 1-6. The output layer is connected to a Softmax activation function to output the predicted probability of each surrounding rock grade. The surrounding rock grade corresponding to the maximum probability is taken as the final predicted surrounding rock grade result. The pre-training process of the deep learning model for surrounding rock grade identification uses historical geological data of completed tunnel projects and corresponding true values of surrounding rock grades to construct an initial training dataset. The sample size of the initial training dataset is >10,000 groups, which are divided into training set, validation set and test set in a ratio of 8:1:1 to complete the training, validation and performance testing of the deep learning model for surrounding rock grade identification.
[0010] Furthermore, in the verification of the prediction results, the true value of the surrounding rock grade is determined comprehensively based on four indicators in the tunnel surrounding rock classification industry: rock mass integrity, rock strength, structural surface development, and groundwater status. During the comparison process, the predicted surrounding rock grade result is compared with the true value of the surrounding rock grade at a construction section every 4-6m. Invalid section data that occurred during construction due to geological disasters or design changes are removed, and only the comparison results of valid construction sections are retained to calculate the prediction accuracy value. The prediction accuracy value is determined by the comprehensive accuracy quantification formula for tunnel surrounding rock grade.
[0011] Furthermore, in the verification of the prediction results, the formula for quantifying the comprehensive accuracy of the tunnel surrounding rock grade is as follows: ,in, The predicted accuracy value for the target tunnel construction section. This represents the total number of valid samples used for accuracy verification within the target tunnel construction section. Valid samples consist of standardized multi-source datasets, predicted surrounding rock grade results, and complete construction section data containing the true values of the surrounding rock grade for the target tunnel construction section. For the index of valid samples, For the first The numerical code of the surrounding rock grade corresponding to each valid sample in the predicted surrounding rock grade result. For the first The surrounding rock grade numeric code corresponding to the true value of the surrounding rock grade for each valid sample, with the surrounding rock grade numeric code corresponding to consecutive integers from 1 to 6 in order from 1 to 6.
[0012] Furthermore, in the closed-loop optimization, the standardized multi-source dataset, the true value of the surrounding rock grade, and the prediction accuracy value are combined to form an incremental training sample set, which is then stored in the incremental sample database. Three types of parallel triggering rules are set for the preset iteration triggering conditions: the first type is a fixed-period trigger, with a triggering period of 30 days. Every 30 days, all incremental training sample sets in the incremental sample database are read, and incremental self-learning training is initiated. The second type is a sample quantity threshold trigger, with a sample quantity threshold of 500 sets. When the cumulative number of incremental training sample sets in the incremental sample database reaches 500 sets, incremental self-learning training is initiated. The third type is an accuracy threshold trigger, with an accuracy threshold of 85%. When the average prediction accuracy value of 20 consecutive construction sections is lower than 85%, incremental self-learning training is initiated. Incremental self-learning training is initiated if any one of the three triggering rules is met. After training, performance is verified using the test set in the initial training dataset. After successful verification, the iterative optimization of the deep learning model for surrounding rock grade identification is completed.
[0013] Furthermore, in the closed-loop optimization, the incremental self-learning training process completes the backpropagation update of the network weights of the deep learning model for rock grade identification through the incremental optimization loss function for rock grade. The formula for the incremental optimization loss function for rock grade is: ,in, This represents the total loss value during incremental self-learning training. This represents the total number of samples in the incremental training sample set. For the index of incremental samples, For the first The numerical code of the surrounding rock grade corresponding to the predicted surrounding rock grade result of each incremental sample. For the first The numerical code of the surrounding rock grade corresponding to the true value of the surrounding rock grade for each incremental sample. This is the function for calculating cross-entropy loss.
[0014] On the other hand, a multi-source information-driven tunnel surrounding rock grade identification system includes: Data acquisition and preprocessing module: acquires design surrounding rock grade data, TSP advanced geological prediction data, high-definition image data of tunnel cross section, and on-site drilling parameter data of the target tunnel construction section, and performs preprocessing on each to obtain a standardized multi-source dataset; Feature extraction and cross-source fusion module: Extracts features from each data point in the standardized multi-source dataset to obtain the corresponding feature vectors, and performs cross-source feature fusion processing on all feature vectors to obtain a global fused feature vector; The grade prediction output module: inputs the global fused feature vector into the pre-trained deep learning model for surrounding rock grade recognition, and outputs the predicted surrounding rock grade of the target tunnel construction section; Prediction result verification module: Obtain the true value of the surrounding rock grade determined by on-site measurement of the target tunnel construction section, compare the predicted surrounding rock grade result with the true value of the surrounding rock grade, calculate and output the prediction accuracy value based on the comparison result; Closed-loop optimization module: The standardized multi-source dataset, the true value of the surrounding rock grade, and the prediction accuracy value are combined to form an incremental training sample set. Based on the preset iteration trigger conditions, the incremental training sample set is input into the deep learning model for surrounding rock grade recognition, and incremental self-learning training is performed to update the network weights of the deep learning model for surrounding rock grade recognition and complete the iterative optimization.
[0015] Compared with existing technologies, this multi-source information-driven method and system for identifying the surrounding rock grade of tunnels has the following advantages: I. This invention combines macroscopic geological background data from the early design stage of tunnels with real-time geological forecasts, on-site borehole data, and tunnel cross-section image data collected during construction through the collaborative application of multi-source tunnel geological data. It constructs feature extraction paths and cross-source fusion mechanisms that match different data types, fully covering all dimensions of information about the tunnel surrounding rock from macroscopic geological background to local real-time status. This eliminates the information limitations caused by a single data source and provides comprehensive information support for the accurate determination of the surrounding rock grade. The overall solution conforms to the entire process of tunnel engineering from survey and design to on-site construction and has high adaptability to engineering implementation.
[0016] Second, this invention constructs a complete closed-loop mechanism for surrounding rock grade prediction, on-site truth value verification, and model iterative optimization. It compares the true value of the surrounding rock grade determined by on-site measurement with the prediction result output by the deep learning model for surrounding rock grade identification. Based on the comparison result, incremental training samples are formed. The incremental self-learning of the deep learning model for surrounding rock grade identification is initiated through preset triggering rules, so as to realize the continuous iterative optimization of the judgment capability of the deep learning model for surrounding rock grade identification. It can continuously adapt to the changes in surrounding rock characteristics under different geological conditions and different construction stages, eliminate the model performance fluctuations under different tunnel engineering scenarios, and ensure the long-term stability and consistency of the surrounding rock grade judgment results.
[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0019] Figure 1 A flowchart of a multi-source information-driven method for identifying the surrounding rock grade of tunnels; Figure 2 A framework diagram of a tunnel surrounding rock grade identification system driven by multi-source information; Figure 3 This is a framework diagram for verifying prediction results in a multi-source information-driven method for identifying the surrounding rock grade of tunnels. Detailed Implementation
[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0021] Example: In the construction scenario of a two-way four-lane highway tunnel project in the mountainous area of central and western China, the tunnel is 2,860m long and has a maximum burial depth of 428m. It traverses Cambrian limestone strata and has adverse geological conditions such as karst and densely jointed zones in some areas. The accurate determination of the surrounding rock grade during construction directly determines the selection of support parameters, the planning of construction procedures, and on-site safety management.
[0022] Data acquisition and preprocessing operations: Through the construction drawing design document issued by the preliminary survey of the tunnel project, the designed surrounding rock grade data of the target construction section is obtained, including the surrounding rock grade classification result given by the design document, corresponding geological lithology, design value of rock mass integrity coefficient, and design value of uniaxial compressive strength of surrounding rock; The TSP203 seismic wave reflection method advanced geological prediction equipment, which is commonly used in tunnel construction period, is used to collect TSP advanced geological prediction data within 100m ahead of the target construction section, including surrounding rock longitudinal wave velocity, spatial position of reflection interface, rock mass integrity parameters, distribution range of adverse geological bodies, and surrounding rock elastic modulus data; A 5-megapixel high-definition industrial camera is used to collect full-range RGB images of the tunnel face and surrounding rock section images before initial support around the tunnel on site after the excavation of the tunnel face is completed and before the initial support construction; Through the measurement while drilling equipment matched with the advanced geological borehole, the on-site drilling parameter data during the construction of the advanced geological borehole is collected in real time, including drilling speed, drill pipe torque, drilling pressure, flushing fluid flow, core recovery rate in the borehole, and recording data of rock mass fragmentation degree; After completing data acquisition, preprocessing is performed on all data: For two types of numerical data, namely designed surrounding rock grade and on-site drilling parameters, the 3σ criterion is used to calculate the data mean and standard deviation, after eliminating outliers exceeding the mean ± 3 times standard deviation, normalization processing is performed; For time-series data such as TSP advanced geological prediction, db4 wavelet is used to perform 3-level decomposition denoising, after removing high-frequency noise components, time-series alignment is completed according to drilling depth, and then normalization processing is performed; For high-definition image data of tunnel sections, the dark channel prior algorithm is used to remove image blur caused by water mist and dust on the tunnel face, after uniformly scaling to 224×224 pixel size, normalization processing of pixel values in the 0-1 interval is performed. Finally, all data are mapped to the 0-1 numerical interval to obtain a standardized multi-source data set, as Figure 1 shown.
[0023] Feature Extraction and Cross-Source Fusion: For different data in standardized multi-source datasets, a matching feature extraction path is adopted. For static numerical data such as design rock grade data, a three-layer fully connected neural network is used for feature extraction. The number of neurons in the three fully connected layers is set to 256, 128, and 64 respectively. Each fully connected layer is followed by a ReLU activation function, ultimately outputting a 64-dimensional fixed-dimensional numerical feature vector. For time-series dynamic data composed of TSP advanced geological prediction data and field borehole parameter data, a three-layer causal convolutional network is first used to extract local time-series features. For sequential features, the kernel size is set to 3 and the stride to 1. The extracted local features are then fed into a two-layer long short-term memory network with a hidden layer dimension of 64 to extract long-distance temporal dependency features, ultimately outputting a 64-dimensional fixed-dimensional temporal feature vector. For high-resolution tunnel cross-section image data, a ResNet18 backbone convolutional neural network is used. After removing the fully connected layers at the network ends used for classification, image feature extraction is performed, ultimately outputting a 512-dimensional fixed-dimensional image feature vector. After completing single-source feature extraction, a global fusion feature vector is constructed using a cross-attention weighted fusion formula for surrounding rock features, such as... Figure 2 As shown; the cross-attention weighted fusion formula for surrounding rock characteristics is: ,in, This represents the global fusion feature vector, where i is the index of the feature vector, taking values of 1, 2, and 3, corresponding to numerical, time-series, and image data, respectively. Let i be the i-th eigenvector. For the fusion weight coefficients corresponding to the i-th feature vector, all The summation result is 1, which is obtained through calculation using the formula: ,in, This serves as the engineering benchmark weighting coefficient, set according to the contribution ratio of various geological exploration methods in the application scenario of tunnel surrounding rock classification engineering. denoted as data validity and completeness, and denoted as the ratio of valid data volume after outlier removal during preprocessing to the total collected data volume. j is the summation variable, corresponding to i.
[0024] Rock grade prediction output: The globally fused feature vector is input into the pre-trained deep learning model for rock grade identification. This deep learning model for rock grade identification adopts a classification network architecture with multi-branch input and single output. The input layer has four parallel input branches, corresponding to the input interfaces of four types of collected data. The intermediate layer has a feature extraction sub-network and a cross-attention fusion sub-network that match the single-source data. The output layer has six output nodes, corresponding to the rock grade of tunnels from level 1 to level 6. The output layer is connected to the Softmax activation function to output the predicted probability corresponding to each rock grade. The rock grade corresponding to the maximum probability is taken as the final predicted rock grade result of the target construction section. In the pre-training stage of this deep learning model for rock grade identification, historical geological data of 30 completed tunnel projects and corresponding ground truth values of rock grades are used to construct an initial training dataset with a sample size of no less than 10,000 sets. The dataset is divided into training set, validation set and test set in a ratio of 8:1:1. The Adam optimizer is used to complete the training, validation and performance testing.
[0025] Prediction Result Verification: Obtain the true value of the surrounding rock grade measured on-site for the target construction section. This true value is determined comprehensively based on four core indicators commonly used in the tunnel surrounding rock classification industry: rock mass integrity, rock strength, structural surface development, and groundwater status. Using a 5m construction section as a unit, compare the predicted surrounding rock grade results of the corresponding sections with the true value of the surrounding rock grade. Invalid section data due to geological disasters or design changes during construction are discarded, retaining only the comparison results of valid construction sections. Using the comprehensive accuracy quantification formula for tunnel surrounding rock grade, calculate and output the prediction accuracy value for the target construction section, such as... Figure 3 As shown; the formula for quantifying the comprehensive accuracy of tunnel surrounding rock grade is: ,in, The predicted accuracy value for the target tunnel construction section. This represents the total number of valid samples used for accuracy verification within the target tunnel construction section. Valid samples consist of standardized multi-source datasets, predicted surrounding rock grade results, and complete construction section data containing the true values of the surrounding rock grade for the target tunnel construction section. For the index of valid samples, For the first The numerical code of the surrounding rock grade corresponding to each valid sample in the predicted surrounding rock grade result. For the first The surrounding rock grade numeric code corresponding to the true value of the surrounding rock grade for each valid sample, with the surrounding rock grade numeric code corresponding to consecutive integers from 1 to 6 in order from 1 to 6.
[0026] Closed-loop optimization: A standardized multi-source dataset, the corresponding ground truth value of the surrounding rock grade, and the prediction accuracy value are combined to form an incremental training sample set, which is then stored in an incremental sample database. Three types of parallel iterative triggering rules are preset: the first is a fixed-period trigger, with a triggering period of 30 days. Every 30 days, all incremental training sample sets in the incremental sample database are read, and incremental self-learning training is initiated. The second is a sample number threshold trigger, with a sample number threshold of 500 sets. When the cumulative number of incremental training sample sets in the incremental sample database reaches 500 sets, incremental self-learning training is initiated. The third is an accuracy threshold trigger, with an accuracy threshold of 85%. When the average prediction accuracy value of 20 consecutive construction sections is lower than 85%, incremental self-learning training is initiated. Incremental self-learning training is initiated if any one of the three triggering rules is met. During training, the backpropagation update of the network weights of the deep learning model for surrounding rock grade identification is completed through the incremental optimization loss function for surrounding rock grade. The formula for the incremental optimization loss function for surrounding rock grade is: ,in, This represents the total loss value during incremental self-learning training. This represents the total number of samples in the incremental training sample set. For the index of incremental samples, For the first The numerical code of the surrounding rock grade corresponding to the predicted surrounding rock grade result of each incremental sample. For the first The numerical code of the surrounding rock grade corresponding to the true value of the surrounding rock grade for each incremental sample. The function for calculating cross-entropy loss is defined. After training, performance verification is performed using the test set reserved in the initial training dataset. Once verification is successful, the deep learning model for identifying surrounding rock grades is iteratively optimized.
[0027] In summary, this embodiment fully implements the entire process of tunnel surrounding rock grade identification driven by multi-source information, closely matches the real construction scenario of tunnel engineering, and achieves accurate and efficient identification of surrounding rock grade through the complementary fusion of multi-source data and a closed-loop self-learning mechanism, providing a feasible and replicable technical solution for tunnel construction safety management.
[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for identifying the surrounding rock grade of tunnels driven by multi-source information, characterized in that, The specific steps of this method are as follows: Data acquisition and preprocessing: Acquire design surrounding rock grade data, TSP advanced geological prediction data, high-definition image data of tunnel cross section, and on-site borehole parameter data of the target tunnel construction section, and preprocess them respectively to obtain standardized multi-source datasets; Feature extraction and cross-source fusion: Features are extracted from each data point in the standardized multi-source dataset to obtain the corresponding feature vectors. Cross-source feature fusion is then performed on all feature vectors to obtain a global fused feature vector. Rock grade prediction output: Input the global fused feature vector into the pre-trained deep learning model for rock grade recognition, and output the predicted rock grade of the target tunnel construction section; Prediction result verification: Obtain the true value of the surrounding rock grade determined by on-site measurement of the target tunnel construction section, compare the predicted surrounding rock grade result with the true value of the surrounding rock grade, calculate and output the prediction accuracy value based on the comparison result; Closed-loop optimization: The standardized multi-source dataset, the true value of the surrounding rock grade, and the prediction accuracy value are combined to form an incremental training sample set. Based on the preset iteration triggering conditions, the incremental training sample set is input into the deep learning model for surrounding rock grade recognition, and incremental self-learning training is performed to update the network weights of the deep learning model for surrounding rock grade recognition, thus completing the iterative optimization.
2. The multi-source information-driven tunnel surrounding rock grade identification method according to claim 1, characterized in that, In the data acquisition and preprocessing process, design surrounding rock grade data is obtained from the construction drawing design documents issued during the early-stage tunnel engineering survey; TSP (Transient Seismic Wave Reflection) advanced geological prediction data is obtained from the advanced geological prediction equipment during tunnel construction; high-definition image data of the tunnel cross-section is acquired on-site using a high-definition industrial camera; and on-site drilling parameter data is acquired in real time using advanced geological borehole measurement and control equipment. The design surrounding rock grade data includes the surrounding rock grade classification results given in the construction drawing design documents, the corresponding geological lithology, the design value of the rock mass integrity coefficient, and the design value of the uniaxial compressive strength of the surrounding rock; the TSP advanced geological prediction data includes the P-wave velocity of the surrounding rock, the spatial location of the reflection interface, rock mass integrity parameters, and adverse geological conditions. The data includes: mass distribution range and surrounding rock elastic modulus data; high-definition tunnel cross-section image data including full-range RGB images of the tunnel face and cross-sectional images of the surrounding rock before the initial support around the tunnel; field drilling parameter data including drilling speed, drill rod torque, drilling pressure, flushing fluid flow rate, core recovery rate, and rock fragmentation degree records of advanced geological boreholes; during preprocessing, outliers were removed from numerical data using the 3σ criterion and then normalized; wavelet transform was used to denoise temporal data and then temporal alignment and normalization were performed; and dark channel prior dehazing was used for image data, followed by size unification and pixel value normalization. All data were mapped to the 0-1 numerical range to obtain a standardized multi-source dataset.
3. The multi-source information-driven tunnel surrounding rock grade identification method according to claim 1, characterized in that, In the feature extraction and cross-source fusion process, different feature extraction paths are adopted for different data in the standardized multi-source dataset. For static numerical data, a three-layer fully connected neural network is used for feature extraction. Each fully connected layer is followed by a ReLU activation function, outputting a fixed-dimensional numerical feature vector. The static numerical data is the design surrounding rock grade data. For time-series dynamic data, a three-layer causal convolutional network is used to extract local time-series features. The extracted local features are then connected to a two-layer long short-term memory network to extract long-distance time-series dependent features, outputting a fixed-dimensional time-series feature vector. The time-series dynamic data is TSP advanced geological prediction data and on-site borehole parameter data. For high-definition image data of tunnel cross sections, a ResNet18 backbone convolutional neural network is used. After removing the classification layer at the end, image feature extraction is performed, outputting a fixed-dimensional image feature vector.
4. The multi-source information-driven tunnel surrounding rock grade identification method according to claim 1, characterized in that, In the feature extraction and cross-source fusion process, the cross-source feature fusion is performed by constructing a global fusion feature vector using a cross-attention weighted fusion formula for surrounding rock features. The cross-attention weighted fusion formula for surrounding rock features is as follows: ,in, This represents the global fusion feature vector, where i is the index of the feature vector, taking values of 1, 2, and 3, corresponding to numerical, time-series, and image data, respectively. Let i be the i-th eigenvector. For the fusion weight coefficients corresponding to the i-th feature vector, all The summation result is 1, which is obtained through calculation using the formula: ,in, This serves as the engineering benchmark weighting coefficient, set according to the contribution ratio of various geological exploration methods in the application scenario of tunnel surrounding rock classification engineering. denoted as data validity and completeness, and denoted as the ratio of valid data volume after outlier removal during preprocessing to the total collected data volume. j is the summation variable, corresponding to i.
5. The multi-source information-driven tunnel surrounding rock grade identification method according to claim 1, characterized in that, In the grade prediction output, the deep learning model for surrounding rock grade identification adopts a classification network architecture with multi-branch input and single output. The input layer has four parallel input branches, corresponding to the input interfaces of four types of data. The intermediate layer has a feature extraction sub-network and a cross-attention fusion sub-network that match the data. The output layer has six output nodes, corresponding to tunnel surrounding rock grades 1-6. The output layer is connected to the Softmax activation function to output the predicted probability of each surrounding rock grade. The surrounding rock grade corresponding to the maximum probability is taken as the final predicted surrounding rock grade result. The pre-training process of the deep learning model for surrounding rock grade identification uses historical geological data of completed tunnel projects and corresponding true values of surrounding rock grades to construct the initial training dataset. The sample size of the initial training dataset is >10,000 groups. It is divided into training set, validation set and test set in a ratio of 8:1:1 to complete the training, validation and performance testing of the deep learning model for surrounding rock grade identification.
6. The multi-source information-driven tunnel surrounding rock grade identification method according to claim 1, characterized in that, In the verification of the prediction results, the true value of the surrounding rock grade is determined by a comprehensive assessment of four indicators in the tunnel surrounding rock classification industry: rock mass integrity, rock strength, structural surface development, and groundwater status. During the comparison process, the predicted surrounding rock grade is compared with the true value of the surrounding rock grade at a construction section every 4-6m. Invalid section data that have been affected by geological disasters or design changes during construction are removed. Only the comparison results of valid construction sections are retained to calculate the prediction accuracy value. The prediction accuracy value is determined by the comprehensive accuracy quantification formula for tunnel surrounding rock grade.
7. The multi-source information-driven tunnel surrounding rock grade identification method according to claim 6, characterized in that, In the verification of the prediction results, the formula for quantifying the comprehensive accuracy of the tunnel surrounding rock grade is as follows: ,in, The predicted accuracy value for the target tunnel construction section. This represents the total number of valid samples used for accuracy verification within the target tunnel construction section. Valid samples consist of standardized multi-source datasets, predicted surrounding rock grade results, and complete construction section data containing the true values of the surrounding rock grade for the target tunnel construction section. For the index of valid samples, For the first The numerical code of the surrounding rock grade corresponding to each valid sample in the predicted surrounding rock grade result. For the first The surrounding rock grade numeric code corresponding to the true value of the surrounding rock grade for each valid sample, with the surrounding rock grade numeric code corresponding to consecutive integers from 1 to 6 in order from 1 to 6.
8. The multi-source information-driven tunnel surrounding rock grade identification method according to claim 1, characterized in that, In the closed-loop optimization, the standardized multi-source dataset, the true value of the surrounding rock grade, and the prediction accuracy value are combined to form an incremental training sample set, which is then stored in the incremental sample database. Three types of parallel triggering rules are set for the preset iteration triggering conditions: the first type is a fixed-period trigger, with a triggering period of 30 days. Every 30 days, all incremental training sample sets in the incremental sample database are read, and incremental self-learning training is initiated. The second type is a sample quantity threshold trigger, with a sample quantity threshold of 500 sets. When the cumulative number of incremental training sample sets in the incremental sample database reaches 500 sets, incremental self-learning training is initiated. The third type is an accuracy threshold trigger, with an accuracy threshold of 85%. When the average prediction accuracy value of 20 consecutive construction sections is lower than 85%, incremental self-learning training is initiated. Incremental self-learning training is initiated if any one of the three triggering rules is met. After training, performance is verified using the test set in the initial training dataset. After successful verification, the iterative optimization of the deep learning model for surrounding rock grade identification is completed.
9. The multi-source information-driven tunnel surrounding rock grade identification method according to claim 8, characterized in that, In the closed-loop optimization, the incremental self-learning training process updates the weights of the deep learning model network for rock grade identification through backpropagation using the incremental optimization loss function for rock grade. The formula for the incremental optimization loss function for rock grade is: ,in, This represents the total loss value during incremental self-learning training. This represents the total number of samples in the incremental training sample set. For the index of incremental samples, For the first The numerical code of the surrounding rock grade corresponding to the predicted surrounding rock grade result of each incremental sample. For the first The numerical code of the surrounding rock grade corresponding to the true value of the surrounding rock grade for each incremental sample. This is the function for calculating cross-entropy loss.
10. A multi-source information-driven tunnel surrounding rock grade identification system, the system being applicable to the multi-source information-driven tunnel surrounding rock grade identification method according to any one of claims 1-9, characterized in that, The system includes: Data acquisition and preprocessing module: acquires design surrounding rock grade data, TSP advanced geological prediction data, high-definition image data of tunnel cross section, and on-site drilling parameter data of the target tunnel construction section, and performs preprocessing on each to obtain a standardized multi-source dataset; Feature extraction and cross-source fusion module: Extracts features from each data point in the standardized multi-source dataset to obtain the corresponding feature vectors, and performs cross-source feature fusion processing on all feature vectors to obtain a global fused feature vector; The grade prediction output module: inputs the global fused feature vector into the pre-trained deep learning model for surrounding rock grade recognition, and outputs the predicted surrounding rock grade of the target tunnel construction section; Prediction result verification module: Obtain the true value of the surrounding rock grade determined by on-site measurement of the target tunnel construction section, compare the predicted surrounding rock grade result with the true value of the surrounding rock grade, calculate and output the prediction accuracy value based on the comparison result; Closed-loop optimization module: The standardized multi-source dataset, the true value of the surrounding rock grade, and the prediction accuracy value are combined to form an incremental training sample set. Based on the preset iteration trigger conditions, the incremental training sample set is input into the deep learning model for surrounding rock grade recognition, and incremental self-learning training is performed to update the network weights of the deep learning model for surrounding rock grade recognition and complete the iterative optimization.