Tunnel surrounding rock quality evaluation index prediction method and system based on attention-lstm
By combining Attention-LSTM model with geophysical exploration and geological sketch data, the problem of low accuracy in predicting the quality of tunnel surrounding rock was solved, achieving high-precision and robust dynamic evaluation and supporting real-time decision optimization for tunnel construction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 中铁科学研究院集团有限公司
- Filing Date
- 2025-04-16
- Publication Date
- 2026-07-31
AI Technical Summary
Existing deep learning-based methods for predicting the quality of tunnel surrounding rock are unable to effectively capture key information under complex geological conditions, resulting in limited evaluation accuracy. In particular, when the physical properties and geological characteristics of the surrounding rock change dynamically during tunnel excavation, it is difficult to accurately extract features.
The Attention-LSTM model is adopted, and the output of LSTM is weighted by the Attention layer. Combined with geophysical information and geological sketch data, a sequence dataset is constructed, and residual connections are used to enhance the model's learning ability, so as to realize the dynamic monitoring and evaluation of the surrounding rock quality.
It improves the prediction accuracy and robustness of the surrounding rock quality assessment model, supports real-time dynamic monitoring, provides timely and accurate information support for tunnel construction, and optimizes data management and model updates.
Smart Images
Figure CN120450508B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of tunnel engineering, specifically to a method and system for predicting tunnel surrounding rock quality evaluation indicators based on Attention-LSTM. Background Technology
[0002] In the field of tunnel engineering, accurately predicting the quality of tunnel surrounding rock is crucial for ensuring construction safety, optimizing design schemes, and reducing project costs. Traditional methods for evaluating surrounding rock quality mainly rely on geological surveys, empirical formulas, and field tests. While these methods can provide basic information about the properties of the surrounding rock to some extent, they often have limitations such as being time-consuming, costly, and highly susceptible to human factors. With the rapid development of artificial intelligence technology, especially the significant advantages of deep learning technology in time series data processing and pattern recognition, new approaches have been provided for evaluating tunnel surrounding rock quality.
[0003] However, existing deep learning-based methods for predicting surrounding rock quality mostly employ standard recurrent neural networks (RNNs) or their variants, such as long short-term memory networks (LSTM). While these models can capture time dependencies when processing long-sequence data, they often struggle to effectively highlight key information, resulting in limited accuracy in evaluating surrounding rock quality under complex geological conditions. Especially during tunnel excavation, the physical properties and geological characteristics of the surrounding rock dynamically change with the progress of excavation. How to accurately extract features crucial for surrounding rock quality evaluation from these sequential data has become an urgent problem to be solved. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for predicting tunnel surrounding rock quality evaluation indicators based on Attention-LSTM, comprising the following steps:
[0005] Step 1: Collect training data and construct a sequence dataset based on the collected training data;
[0006] Step 2: Based on the LSTM network, the Attention layer is used to weight the output of the LSTM to construct an Attention-LSTM model;
[0007] Step 3: Train the Attention-LSTM model based on the constructed sequence dataset to obtain the trained Attention-LSTM model;
[0008] Step 4: Collect sample point data and input it into the trained Attention-LSTM model to obtain the probability distribution of the hardness, water flow status, and integrity of the sample points.
[0009] Furthermore, the process of collecting training data and constructing a sequence dataset based on the collected training data includes:
[0010] The rock physical parameters of the surrounding rock are obtained by using elastic wave geophysical exploration equipment to detect the area in front of the tunnel working face, and a set of data is obtained at each set distance.
[0011] Based on the geological sketch data of the tunnel working face, data on the hardness and integrity of the surrounding rock and the state of groundwater are extracted, and a set of data is obtained at each set extraction distance;
[0012] The training samples are divided into segments with sequential relationships based on the mileage prefix, and samples with a difference of less than or equal to a set distance are considered as the same segment. Each training sample consists of geophysical data of partially excavated and unexcavated segments, geological features of the excavated segments, and real labels.
[0013] Furthermore, the aforementioned LSTM-based network, which weights the LSTM output through an Attention layer to construct an Attention-LSTM model, includes:
[0014] a. Encoding stage: The input sequence is encoded through one or more LSTM layers, and a hidden state htht is generated at each time step;
[0015] b. Attention Calculation: In each decoding step, based on the current decoder state s t-1 and the hidden states of all encoders {h1,h2,...,h T}, α is obtained by normalization using the softmax function. i According to attention weight α i We obtain the context vector C by weighted summation of the encoder's hidden states. t ;
[0016] c. Decoding stage: The context vector C is... t Combined with the output of the previous decoder, it serves as the input for the current decoding step, and this process is repeated until the entire output sequence is generated.
[0017] Furthermore, the collected sample point data is input into the trained Attention-LSTM model to obtain the probability distributions of the hardness, water inrush state, and integrity corresponding to the sample points, including:
[0018] Collect and store geophysical exploration data and tunnel face data; remove outliers; perform feature engineering to construct the metrics required for model learning; divide the data sequence and perform Z-Score normalization; input the trained Attention-LSTM model for prediction, and output the probability distribution of hardness, water inrush state, and integrity corresponding to each sample point; excavate for verification and add cases, add the real data after excavation verification to the database, and repeat the process to improve accuracy.
[0019] Furthermore, the geophysical information data includes P-wave velocity, S-wave velocity, density, and Poisson's ratio.
[0020] Furthermore, the hardness of the geological sketch data is divided into extremely hard rock, hard rock, relatively soft rock, soft rock, and extremely soft rock; the integrity is divided into complete, relatively complete, relatively broken, broken, and extremely broken; and the groundwater state is divided into dry, wet or dripping water, rain-like or linear flow water, and gushing water.
[0021] Furthermore, the Attention-LSTM model also includes residual connections, which are added after the LSTM layer in the decoding stage.
[0022] The tunnel surrounding rock quality evaluation index prediction system based on Attention-LSTM, applying the aforementioned method for predicting tunnel surrounding rock quality evaluation indexes based on Attention-LSTM, includes: a data collection module, a data preprocessing module, a model training module, a prediction module, a database module, and a data processing module; the data collection module, data preprocessing module, model training module, prediction module, and database module are respectively connected to the data processing module;
[0023] The data collection module is used to collect geophysical information data and geological sketch data.
[0024] The data preprocessing module is used to construct sequence datasets, remove outliers, perform feature engineering, and normalize data.
[0025] The model training module is used to train the model using the Attention-LSTM algorithm;
[0026] The prediction module is used to make predictions based on the input data and output the probability distribution of the surrounding rock quality evaluation index.
[0027] The database module is used to store training data and real data after excavation and verification, in order to support the continuous optimization and updating of the model.
[0028] The beneficial effects of this invention are: improved prediction accuracy: by weighting the hidden state of the LSTM output through the Attention layer, the model can pay more attention to the features and time steps that have an important impact on the evaluation of surrounding rock quality, effectively improving the prediction accuracy of key indicators such as hardness, water inrush state, and integrity.
[0029] Enhancing Model Robustness: This invention constructs a sequence dataset containing geophysical information data and geological sketch data, and combines residual connections to enhance the model's learning ability, enabling the model to exhibit stronger robustness and generalization ability when facing complex and variable geological conditions.
[0030] Achieving dynamic prediction: The method supports real-time data updates as the tunnel excavation progresses. Through continuous data input and model prediction, it enables dynamic monitoring and evaluation of the surrounding rock quality, providing timely and accurate information support for construction decisions.
[0031] Optimized data management: By integrating data collection, preprocessing, model training, prediction, and database modules, a complete tunnel surrounding rock quality evaluation system was constructed, realizing systematic management and efficient utilization of data, and facilitating continuous optimization and updating of the model. Attached Figure Description
[0032] Figure 1 This is a flowchart illustrating the prediction method for tunnel surrounding rock quality evaluation indicators based on Attention-LSTM.
[0033] Figure 2 This is a schematic diagram illustrating the principle of a tunnel surrounding rock quality evaluation index prediction system based on Attention-LSTM.
[0034] Figure 3 This is a schematic diagram of the Attention-LSTM structure. Detailed Implementation
[0035] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.
[0036] The features and performance of the present invention will be further described in detail below with reference to embodiments.
[0037] like Figure 1 As shown, the method for predicting tunnel surrounding rock quality evaluation indicators based on Attention-LSTM includes the following steps:
[0038] Step 1: Collect training data and construct a sequence dataset based on the collected training data;
[0039] Step 2: Based on the LSTM network, the Attention layer is used to weight the output of the LSTM to construct an Attention-LSTM model;
[0040] Step 3: Train the Attention-LSTM model based on the constructed sequence dataset to obtain the trained Attention-LSTM model;
[0041] Step 4: Collect sample point data and input it into the trained Attention-LSTM model to obtain the probability distribution of the hardness, water flow status, and integrity of the sample points.
[0042] The process of collecting training data and constructing a sequence dataset based on the collected training data includes:
[0043] The rock physical parameters of the surrounding rock are obtained by using elastic wave geophysical exploration equipment to detect the area in front of the tunnel working face, and a set of data is obtained at each set distance.
[0044] Based on the geological sketch data of the tunnel working face, data on the hardness and integrity of the surrounding rock and the state of groundwater are extracted, and a set of data is obtained at each set extraction distance;
[0045] The training samples are divided into segments with sequential relationships based on the mileage prefix, and samples with a difference of less than or equal to a set distance are considered as the same segment. Each training sample consists of geophysical data of partially excavated and unexcavated segments, geological features of the excavated segments, and real labels.
[0046] The aforementioned LSTM-based network, which uses an Attention layer to weight the LSTM outputs to construct an Attention-LSTM model, includes:
[0047] a. Encoding stage: The input sequence is encoded through one or more LSTM layers, and a hidden state htht is generated at each time step;
[0048] b. Attention Calculation: In each decoding step, based on the current decoder state s t-1 and the hidden states of all encoders {h1,h2,...,h T}, α is obtained by normalization using the softmax function. i According to attention weight α i We obtain the context vector C by weighted summation of the encoder's hidden states. t ;
[0049] c. Decoding stage: The context vector C is... t Combined with the output of the previous decoder, it serves as the input for the current decoding step, and this process is repeated until the entire output sequence is generated.
[0050] The collected sample point data is input into the trained Attention-LSTM model to obtain the probability distributions of the hardness, water inrush state, and integrity of the sample points, including:
[0051] Collect and store geophysical exploration data and tunnel face data; remove outliers; perform feature engineering to construct the metrics required for model learning; divide the data sequence and perform Z-Score normalization; input the trained Attention-LSTM model for prediction, and output the probability distribution of hardness, water inrush state, and integrity corresponding to each sample point; excavate for verification and add cases, add the real data after excavation verification to the database, and repeat the process to improve accuracy.
[0052] The geophysical information data includes P-wave velocity, S-wave velocity, density, and Poisson's ratio.
[0053] The hardness of the geological sketch data is divided into extremely hard rock, hard rock, relatively soft rock, soft rock, and extremely soft rock; the integrity is divided into complete, relatively complete, relatively broken, broken, and extremely broken; and the groundwater state is divided into dry, wet or dripping water, rain-like or linear water flow, and gushing water.
[0054] The Attention-LSTM model also includes residual connections, which are added after the LSTM layer in the decoding stage.
[0055] like Figure 2 As shown, the tunnel surrounding rock quality evaluation index prediction system based on Attention-LSTM, which applies the aforementioned Attention-LSTM-based tunnel surrounding rock quality evaluation index prediction method, includes: a data collection module, a data preprocessing module, a model training module, a prediction module, a database module, and a data processing module; the data collection module, data preprocessing module, model training module, prediction module, and database module are respectively connected to the data processing module;
[0056] The data collection module is used to collect geophysical information data and geological sketch data.
[0057] The data preprocessing module is used to construct sequence datasets, remove outliers, perform feature engineering, and normalize data.
[0058] The model training module is used to train the model using the Attention-LSTM algorithm;
[0059] The prediction module is used to make predictions based on the input data and output the probability distribution of the surrounding rock quality evaluation index.
[0060] The database module is used to store training data and real data after excavation and verification, in order to support the continuous optimization and updating of the model.
[0061] Specifically, including:
[0062] I. Model Training
[0063] 1. Training data includes
[0064] (1) Geophysical Information Data: Using elastic wave geophysical exploration equipment (including but not limited to TSP, HSP, SAP, etc.), the area in front of the tunnel working face is explored to obtain the surrounding rock physical parameters, including P-wave velocity, S-wave velocity, density, and Poisson's ratio. A set of data can be obtained every 120m.
[0065] (2) Using geological sketch data of the tunnel working face, extract data on the hardness and integrity of the surrounding rock and the state of groundwater. A set of data can be obtained every 3 meters.
[0066] (3) Hardness is classified into: extremely hard rock, hard rock, relatively soft rock, soft rock, and extremely soft rock.
[0067] The degree of integrity is divided into: complete, relatively complete, relatively fragmented, fragmented, and extremely fragmented.
[0068] Groundwater conditions are categorized as follows: dry, wet or dripping, rain-like or linear flow, and gushing.
[0069] 2. Model training using the Attention-LSTM machine learning algorithm.
[0070] ① Sequence Dataset Construction
[0071] a. Segment division with sequential relationships. Geological features exhibit strong spatial continuity and autocorrelation; therefore, multiple adjacent geophysical data points are used to form a sample. Since the data consists of multiple discontinuous tunnel segments, it is necessary to first divide the data into segments with sequential relationships. For each mileage prefix, samples within 100m of each other are considered as the same segment.
[0072] b. Training dataset structure. Each training sample consists of three parts: geophysical data of partially excavated and unexcavated sections, geological features of the excavated sections, and ground truth labels.
[0073] ②Attention-LSTM network
[0074] LSTM (Long Short-Term Memory) is a special type of recurrent neural network (RNN) architecture used to process and predict important events with very long intervals and delays in time series data. The attention layer weights the output of the LSTM, enhancing the representational power of local features. Finally, a fully connected layer with residual connections and a softmax function are used to obtain the probability of each class. The overall structure is as follows... Figure 3 As shown.
[0075] a. Encoding stage: The input sequence is encoded through one or more LSTM layers, and a hidden state htht is generated at each time step. These hidden states together constitute the encoded representation of the input sequence.
[0076] b. Attention Calculation: In each decoding step, based on the current decoder state s t-1 and the hidden states of all encoders {h1,h2,...,h T}, calculate attention weight α i The calculation method usually involves using s t-1 and each h i The data is fed into a feedforward neural network and then normalized using the softmax function to obtain α. i Using these weights, the hidden states of the encoder are summed to obtain the context vector C. t .
[0077] c. Decoding stage: The context vector C is... t The output of the previous decoder is combined with the current input and used as the input for the current decoding step. The decoder is also an LSTM that generates the output for the next time step based on the current input and the previous decoding state. This process is repeated until the entire output sequence is generated.
[0078] d. Residual Connections: Residual connections are a technique in deep learning models that increases model depth while maintaining or improving performance. In Attention-LSTM, applying residual connections typically means adding the input of one layer directly to the output of the next layer. For LSTM models with attention mechanisms, residual connections can be added after the LSTM layers in the decoding stage, meaning the decoder's input can be directly added to the decoder's output, or the output of one layer can be added to the output of the next layer. This helps information propagate forward more effectively and reduces the vanishing gradient problem during training.
[0079] II. Intelligent Prediction Steps Using Elastic Wave Method
[0080] 1. Overall forecasting approach:
[0081] (1) According to the time series, geophysical information and geological information can be input simultaneously or at different times. Each time information is input, the program runs once and outputs a prediction result, namely the integrity, hardness, and probability of each type of groundwater.
[0082] (2) For example, if the geological information (3 meters) and geophysical information (120 meters) of the tunnel face are entered for the first time, the first prediction result is obtained. After the tunnel is excavated for 3 meters for the second time, the tunnel face information is entered again, and the process is repeated, outputting the surrounding rock result 117 meters ahead.
[0083] (3) And so on, infinitely approaching the true value.
[0084] 2. Specific steps:
[0085] ① Geophysical data and tunnel face data obtained from a certain geophysical test
[0086] According to the proposed data entry specifications, the data to be entered into the database is a three-dimensional data volume (.CSV format). Based on the xyz coordinates and mining direction in the data, the location of each sample point and the geophysical data indicators can be determined.
[0087] ② Remove outliers
[0088] The data may contain samples with missing geophysical indicators or those with significantly deviated values. Samples with abnormal geophysical indicators are usually removed.
[0089] ③ Feature Engineering
[0090] The original data only contained four columns of indicators: P-wave velocity, P-wave to S-wave velocity ratio, density, and Poisson's ratio. To facilitate model learning, indicators were constructed by taking the logarithm of the skewed data and using feature derivation.
[0091] ④ Neural network prediction
[0092] Before training, the data sequence is first divided into multiple consecutive segments with sequential relationships based on mileage prefixes and mileage numbers. Next, Z-score normalization is applied to the data to prevent data dimensions from affecting neural network convergence. Finally, the data is input into the trained neural network for prediction, outputting the probability distribution of hardness, water inrush status, and integrity for each sample point.
[0093] ⑤ Excavate and verify, then add case studies.
[0094] Subsequently, the actual hardness, water inflow status, and integrity after excavation verification are continuously added to the database, and the process is repeated to improve the accuracy of the module.
[0095] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
Claims
1. A method for predicting tunnel surrounding rock quality evaluation indicators based on Attention-LSTM, characterized in that, Includes the following steps: Step 1: Collect training data and construct a sequence dataset based on the collected training data; Step 2: Based on the LSTM network, the Attention layer is used to weight the output of the LSTM to construct an Attention-LSTM model; Step 3: Train the Attention-LSTM model based on the constructed sequence dataset to obtain the trained Attention-LSTM model; Step 4: Collect sample point data and input it into the trained Attention-LSTM model to obtain the probability distribution of hardness, water flow status, and integrity of the sample points. The process of collecting training data and constructing a sequence dataset based on the collected training data includes: The rock mass profile is obtained by probing the area in front of the tunnel working face using elastic wave geophysical equipment. A set of geological sketch data of the tunnel working face is obtained at each set distance. Based on the geological sketch data of the tunnel working face, information on the hardness, integrity, and groundwater status of the surrounding rock is extracted. The training samples are divided into segments with sequential relationships based on the mileage prefix, and samples with a difference of less than or equal to a set distance are considered as the same segment. Each training sample consists of geophysical data of some excavated segments and unexcavated segments, geological features of the excavated segments, and real labels. The aforementioned LSTM-based network, which uses an Attention layer to weight the LSTM outputs to construct an Attention-LSTM model, includes: a. Encoding stage: The input sequence is encoded through one or more LSTM layers, generating a hidden state h at each time step. T ; b. Attention calculation: In each decoding step, based on the current decoder state... and the hidden state of all encoders ,pass Function normalization yields Based on attention weights The context vector is obtained by weighted summation of the encoder's hidden states. ; c. Decoding phase: The context vector is... Combined with the output of the previous decoder, it serves as the input for the current decoding step, and this process is repeated until the entire output sequence is generated.
2. The method for predicting tunnel surrounding rock quality evaluation indicators based on Attention-LSTM according to claim 1, characterized in that, The collected sample point data is input into the trained Attention-LSTM model to obtain the probability distributions of the hardness, water inrush state, and integrity of the sample points, including: Collect geophysical and tunnel face data and remove outliers; divide the data sequence and perform Z-Score normalization; input the trained Attention-LSTM model for prediction and output the probability distribution of surrounding rock hardness, water inrush state, and integrity for each sample point; excavate for verification and add cases, add the real data after excavation verification to the database, and repeat the process to improve accuracy.
3. The method for predicting tunnel surrounding rock quality evaluation indicators based on Attention-LSTM according to claim 2, characterized in that, The geophysical data includes P-wave velocity, S-wave velocity, density, and Poisson's ratio.
4. The method for predicting tunnel surrounding rock quality evaluation indicators based on Attention-LSTM according to claim 1, characterized in that, The hardness of the geological sketch data is divided into extremely hard rock, hard rock, relatively soft rock, soft rock, and extremely soft rock; the integrity is divided into complete, relatively complete, relatively broken, broken, and extremely broken; and the groundwater state is divided into dry, wet or dripping water, rain-like or linear water flow, and gushing water.
5. The method for predicting tunnel surrounding rock quality evaluation indicators based on Attention-LSTM according to claim 1, characterized in that, The Attention-LSTM model also includes residual connections, which are added after the LSTM layer in the decoding stage.
6. A tunnel surrounding rock quality evaluation index prediction system based on Attention-LSTM, characterized in that, The method for predicting tunnel surrounding rock quality evaluation indicators based on Attention-LSTM according to any one of claims 1-5 includes: a data collection module, a data preprocessing module, a model training module, a prediction module, a database module, and a data processing module; wherein the data collection module, data preprocessing module, model training module, prediction module, and database module are respectively connected to the data processing module; The data collection module is used to collect geophysical information data and geological sketch data. The data preprocessing module is used to construct sequence datasets, remove outliers, perform feature engineering, and normalize data. The model training module is used to train the model using the Attention-LSTM algorithm; The prediction module is used to make predictions based on the input data and output the probability distribution of the surrounding rock quality evaluation index. The database module is used to store training data and real data after excavation and verification, in order to support the continuous optimization and updating of the model.