Tunnel disaster identification method based on geological inference process
Through the method based on the geological inference process, Bayesian model and iterative training technology, the problem of difficult prediction of rock burst disasters in tunnel construction is solved, and the accurate inference of geological data in front of the tunnel and rock burst level prediction are achieved, and the tunnel is made by using Bayesian model and iterative training technology. Construction provides safety guarantees.
Patent Information
- Application Number
- CN202411570325.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-06
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2044-11-06
AI Technical Summary
It is difficult to predict rock burst disasters during existing tunnel construction, the existing TSP method is troublesome to operate, and the detection distance is limited, making it difficult to effectively predict geological parameters and rock burst risks in front of the tunnel.
Using the tunnel disaster determination method based on the geological inference process, the rock explosion disaster is constructed by collecting the unexcavated TSP data in front of the palm face and the excavated geological data behind the palm face, and using Bayesian inference model and iterative training methods. The model is judged to realize the inference of unexcavated geological data and prediction of rock burst ratings.
Accurate inference of the unexcavated part of the geological data in front of the tunnel and prediction of rock burst disaster levels, providing important data support, and providing guarantees for the safety of tunnel geological construction.
Smart Images

Figure CN119067177A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of tunnel surrounding rock detection, and in particular to a tunnel disaster identification method based on geological inference process. Background Art
[0002] During the tunnel construction process, the tunnel engineering disasters caused often have a great impact on the tunnel construction, which may slow down the progress of tunnel construction or endanger the safety of construction workers. Therefore, it is very important to predict and identify tunnel geological disasters in advance. This method mainly focuses on the inference of geological parameters of the unexcavated area in front of the tunnel face during tunnel construction and the prediction and identification function of rock burst.
[0003] Rockburst refers to a catastrophic phenomenon in which rock suddenly releases energy due to the concentration of ground stress during underground engineering and mining. Its main characteristics include the sudden fracture and fragmentation of rocks, the generation of shock waves, and the sudden eruption and collapse of rocks or ore bodies. The main reason for its occurrence is high ground stress. Under high ground stress and hard rock environment, the stress of the surrounding rock is redistributed after the disturbance of tunnel excavation. When the stress borne by the surrounding rock reaches the limit, it may lead to the occurrence of rockburst disaster.
[0004] The main influencing factors of rockburst tunnel disasters include the magnitude of geostress in the tunnel area and the rock mass strength of the surrounding rock. Among them, the rock mass strength of the surrounding rock depends on the rock hardness of the rock contained in the surrounding rock itself, the groundwater conditions of the surrounding rock, and the integrity of the surrounding rock. When constructing the tunnel surrounding rock, it is necessary to detect the surrounding rock and collect the basic properties of the surrounding rock to provide data support for the later construction. However, the existing TSP method is cumbersome to operate and the detection distance is limited. Therefore, a tunnel disaster identification method based on the geological inference process is proposed. Summary of the invention
[0005] In order to solve the technical problems existing in the prior art, the present invention provides a tunnel disaster identification method based on geological inference process.
[0006] The present invention is implemented by the following technical solution: a tunnel disaster identification method based on geological inference process, comprising the following steps: Step S1, collecting unexcavated TSP data in front of the tunnel face and excavated geological data behind the tunnel face; Step S2, using the collected unexcavated TSP data in front of the tunnel face and the excavated geological data behind the tunnel face to infer the unexcavated geological data in front of the tunnel face using a geological parameter inference model; Step S3, constructing a rockburst disaster identification model; Step S4, pre-processing the unexcavated TSP data in front of the tunnel face and the inferred unexcavated geological data in front of the tunnel face to form a data set; Step S5: Model training, optimization and evaluation: The model is trained and optimized using an iterative training method, which can be specifically divided into initial model training and iterative training; Use the test set for evaluation, calculate the classification accuracy, predict and output the true category and predicted category of the test set, and select accuracy as the performance indicator; Step S6: Using the trained and optimized model to identify rockburst disasters.
[0007] As a further improvement of the above scheme, the unexcavated TSP data in front of the tunnel face in step S1 include longitudinal wave velocity, shear wave velocity, Poisson's ratio, density, static Young's modulus, dynamic Young's modulus and direct wave velocity; the excavated geological data behind the tunnel face include rock hardness, groundwater conditions, integrity and ground stress.
[0008] As a further improvement of the above solution, the Bayesian inference model in step S2 includes:
[0009] in, : Posterior probability, the probability that the parameter θ is the true value after given the data D; : Likelihood function, which indicates the probability of data D appearing when parameter θ is true; : Prior probability refers to the original belief about parameter θ before observing the current data; : The marginal likelihood, also called the normalizing constant or evidence factor, is the overall probability of the observation D occurring.
[0010] As a further improvement of the above scheme, the rock burst disaster identification model in step S3 includes an input layer, a convolutional layer 1, a convolutional layer 2, a pooling layer, a Flatten layer, a fully connected layer and an output layer which are connected in sequence.
[0011] As a further improvement of the above solution, the preprocessing in step S4 includes the following steps: Step A1: normalize the unexcavated TSP data features in front of the tunnel face: The MinMaxScaler function is used to scale the numerical features in the unexcavated TSP data in front of the tunnel face to the interval [0,1]; Step A2: Encode the inferred unexcavated geological data and disaster categories ahead of the tunnel face: The text features in the inferred unexcavated geological data ahead of the tunnel face are converted into numerical values by mapping; Step A3: Map target labels: Map the string "rockburst level" to a numeric label; Step A4: Split training set and test set: The dataset is divided into training set and test set in a ratio of 80:20; Step A5, one-hot encoding: OneHotEncoder is applied to the target variable, i.e., rockburst grade, to convert the category labels into one-hot vectors.
[0012] As a further improvement of the above solution, the model training and optimization in step S5 includes the following steps: Step B1: Model compilation: The Adam optimizer performs initial configuration on the model, allowing the model to optimize parameters using an adaptive learning rate. The loss function is categorical_crossentropy, the goal is to minimize the error between the category output probability and the actual label; Step B2, callback function: Use ModelCheckpoint and EarlyStopping to help adjust the model, automatically save the best model, prevent overfitting and improve training efficiency; Step B3: Training model: The training set is input into the model for parameter optimization, and the performance is evaluated and verified after multiple iterations.
[0013] As a further improvement of the above solution, the model evaluation in step S5 includes the following steps: Step C1: Visualization of model training results: Use Matplotlib to plot the loss and accuracy curves during the training and validation phases; Step C2, model evaluation: Evaluate on the test set, calculate the classification accuracy, predict and output the true category and predicted category of the test set, and select accuracy as the performance indicator.
[0014] As a further improvement of the above scheme, the iterative evaluation includes 100-250 training iterations, using 10-20 samples in each iteration, and splitting 20-30% of the data from the training data for verification to evaluate the performance of the model during the training process.
[0015] As a further improvement of the above scheme, the unexcavated TSP data features in front of the tunnel face are normalized using the MinMaxScaler function, where the MinMaxScaler formula is:
[0016] in is the original data, and are the minimum and maximum values of the feature, respectively.
[0017] Compared with the prior art, the present invention has the following beneficial effects: After completing the statistics of the prior probability and the conditional probability, the present invention uses the method of using the prior data as the prior data of the next step. Each time the observed D1, D2, and D3 states are input, the prior probability can be combined with the influence of the three physical states to generate the posterior probability. The generated posterior probability is then used as the prior probability. By inputting the observed D1, D2, and D3 states multiple times, the prior probability can be updated multiple times, thereby realizing a dynamic update process of the prior probability combined with the influence of the observed state. According to the posterior probability distribution after each update, the state with the highest probability in the rock hardness state can be selected as the inference result. The geological data of the excavated tunnel is processed as a priori, that is, a general judgment of the rear of the tunnel. Then, TSP advanced geophysical exploration is used to collect TSP data behind the face, and the collected TSP data is input into the inference model as observation data. The input observation data will affect the priori according to the conditional probability relationship between TSP data and geological data to obtain the posterior, and the posterior is the geological data inference result of the unexcavated section corresponding to the input TSP data; then the updated posterior is used as the priori, and the subsequent TSP data is continuously input, so as to continuously update the posterior to obtain the geological data inference result, and realize the inference of the four indicators of rock hardness, groundwater conditions, integrity and ground stress of the geological data of the unexcavated part; The rockburst hazard identification model is used in combination with the collected excavated geological data behind the tunnel face and the inferred unexcavated geological data in front of the tunnel face to realize the rockburst hazard level prediction operation of the unexcavated part in front of the tunnel face, providing data support for tunnel geological construction. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 A flow chart of a tunnel disaster identification method based on geological inference process provided by the present invention; Figure 2 The inference model provided by the present invention is used to infer the principle diagram of the unexcavated geological data in front of the tunnel face; Figure 3 A schematic diagram of the structure of the rockburst disaster identification model provided by the present invention; Figure 4 A verification diagram of iterative training of the rockburst disaster identification model provided by the present invention; DETAILED DESCRIPTION
[0019] The present invention is further described below in conjunction with the accompanying drawings and specific implementation methods. It should be noted that, under the premise of no conflict, the various embodiments or technical features described below can be arbitrarily combined to form a new embodiment.
[0020] Example 1
[0021] Please combine Figure 1-Figure 4 , a tunnel disaster identification method based on geological inference process of this embodiment includes the following steps: Step S1, collecting unexcavated TSP data in front of the tunnel face and excavated geological data behind the tunnel face; Step S2, using the collected unexcavated TSP data in front of the tunnel face and the excavated geological data behind the tunnel face to infer the unexcavated geological data in front of the tunnel face using a geological parameter inference model; Step S3, constructing a rockburst disaster identification model; Step S4, pre-processing the unexcavated TSP data in front of the tunnel face and the inferred unexcavated geological data in front of the tunnel face to form a data set; Step S5: Model training, optimization and evaluation: The model is trained and optimized using an iterative training method, which can be specifically divided into initial model training and iterative training; Use the test set for evaluation, calculate the classification accuracy, predict and output the true category and predicted category of the test set, and select accuracy as the performance indicator; Step S6: Using the trained and optimized model to identify rockburst disasters.
[0022] Example 2
[0023] A geological inference method for tunnel disaster identification based on a geological inference process comprises the following steps: Step S11, collecting unexcavated TSP data in front of the tunnel face and excavated geological data behind the tunnel face; Collect the unexcavated TSP data in front of the tunnel face and the excavated geological data behind the tunnel face. The unexcavated TSP data in front of the tunnel face include longitudinal wave velocity, shear wave velocity, Poisson's ratio, density, static Young's modulus, dynamic Young's modulus and direct wave velocity; the excavated geological data behind the tunnel face include rock hardness, groundwater conditions, integrity and ground stress; The basic principle of TSP method, i.e. earthquake prediction or advanced geological prediction ahead of the tunnel, is as follows: Excitation holes are arranged within a certain range of the side wall near the tunnel face. Seismic waves are artificially excited in the holes. The generated seismic waves propagate in the tunnel surrounding rock in the form of spherical waves. When the surrounding rock wave impedance changes (for example, when encountering karst, fault or rock layer interface), part of the seismic waves will be reflected back, and the other part will continue to propagate forward; the reflected seismic waves are received by high-precision receivers and transmitted to the host to form seismic wave records. The returned seismic wave records can be used to calculate the longitudinal wave velocity, shear wave velocity, direct wave velocity, etc. within the coverage range of TSP advanced geological forecasting; based on the returned seismic wave information, the geological conditions in front of the tunnel face can be judged; Step S12, dividing the collected data into data sets; The data set is divided based on the excavated geological data parameters behind the tunnel face and the unexcavated TSP data in front of the tunnel face; Among them, the rock hardness, groundwater conditions, integrity and ground stress in the excavated geological data parameters behind the tunnel face are used as references, and the data sets are divided into three groups: D1, D2 and D3. The selection of the types of unexcavated TSP data in front of the tunnel face corresponding to the excavated geological data parameter indicators behind the tunnel face is as follows: Rock hardness: D1: dynamic Young's modulus, D2: longitudinal wave velocity, D3: density; Groundwater conditions: D1: transverse wave velocity, D2: longitudinal wave velocity, D3: direct wave velocity; Completeness: D1: longitudinal wave velocity, D2: Poisson's ratio, D3: static Young's modulus; Geostress: D1: direct wave velocity, D2: Poisson's ratio, D3: density; Step S13, performing conditional probability calculation on the data set to obtain conditional probability; Conditional probability statistics (taking the conditional probability of calculating rock hardness as an example): Regarding the conditional probability (likelihood probability) of rock hardness, three types of unexcavated TSP data in front of the tunnel face related to rock hardness are considered, including the dynamic Young's modulus, longitudinal wave velocity, and Poisson's ratio of the unexcavated TSP data in front of the tunnel face. When considering the conditional probability distribution between the three physical properties and rock hardness, the following steps are used to determine the conditional probability: (1) Data collection and status definition: Data Collection: TSP observational data including rock hardness and corresponding dynamic Young's modulus, longitudinal wave velocity and Poisson's ratio are collected.
[0024] State definition: define the state for each variable; for example, the dynamic Young's modulus can be divided into five states according to the specific numerical range of the collected data, which can be briefly described as low, relatively low, medium, relatively high, and high, to facilitate the statistics of conditional probability.
[0025] (2) Frequency analysis and frequency calculation: Frequency distribution: For each rock hardness state, the frequencies of different combinations of the three attributes are counted, such as the number of occurrences of dynamic Young's modulus under different rock hardness states and the number of occurrences of different rock hardness states.
[0026] Ratio conversion: Convert the joint occurrence frequency into probability. The conditional probability is expressed as the probability of other attribute states appearing under different states of one attribute. The probability calculation formula is as follows:
[0027] By calculating the conditional probability, the conditional probability matrix of rock hardness and dynamic Young's modulus, longitudinal wave velocity, and Poisson's ratio can be generated. An example of conditional probability is shown in Table 1: Table 1 is the conditional probability table of rock hardness and dynamic Young's modulus:
[0028] Step S14, using the Bayesian inference model to calculate the prior probability of the data set to obtain the prior probability; Step S15, using the Bayesian inference model to perform a first posterior probability calculation on the data set to obtain a first posterior probability; Step S16, taking the first posterior probability as the prior probability, and inputting the data set into the Bayesian inference model to perform the second posterior probability; Step S17, repeating steps S15 and S16 to perform multiple calculations using the Bayesian inference model data set; Step S18, selecting the state with the highest probability as the inference result according to the content calculated in step S17; Probability update process (taking rock hardness as an example): (1) Confirmation of prior probability The prior probability is confirmed by known information. Prior probability is an important concept in the Bayesian inference process. It expresses the view of a thing in the form of probability. At this time, according to the excavated geological data parameter indicators behind the tunnel face collected in the data sets D1, D2 and D3, the prior probability calculation of the excavated geological data parameter indicators behind the tunnel face is realized, and the probability of rock hardness under different hardness conditions is calculated, so as to obtain the prior probability of rock hardness. ; The specific calculation method is to statistically analyze the geological parameters of the excavated section behind the tunnel face. Taking rock hardness as an example, the rock hardness data within 50m behind the tunnel face is recorded and statistically analyzed. One piece of data is recorded every 2m. The frequency of occurrence of each rock hardness is recorded, and the probability of occurrence of each rock hardness is calculated. The calculation method is as follows:
[0029] After the calculation is completed, some rock hardness states may not have appeared, resulting in a calculated probability of 0. When the prior probability is 0, the posterior probability will always be updated to 0, which does not meet the inference requirements. Therefore, the rock hardness state with a probability of zero needs to be adjusted. Increase the probability of all rock hardness states with a probability of 0 by 2%, and redistribute them according to the increased probability distribution so that the sum of the rock hardness probabilities is 1, so that it meets the probability update requirements; When the prior probability of rock hardness is calculated to be: P(hard rock)=0.2, P(harder rock)=0.4, P(softer rock)=0.1, P(soft rock)=0.2, P(extremely soft rock)=0.1, it means that under the current circumstances, the probability of harder rock appearing is the highest, while the probability of extremely soft rock appearing is the lowest.
[0030] (2) Conditional probability calculation: The conditional probability calculation of rock hardness is realized based on the collected data sets D1, D2 and D3, so as to obtain the conditional probability matrix value, which also represents the probability of observing a specific dynamic Young's modulus state under a certain rock hardness.
[0031] (3) Calculation of evidence factor: The evidence factor is calculated by combining the likelihood of all hardness states, i.e. the marginal probability of the dynamic Young's modulus in all hardness states. If the state of the observed Young's modulus is low, the calculation method is as follows:
[0032] The evidence factor is a normalization constant used to ensure that the sum of the posterior probabilities is 1. By calculating the likelihood of the dynamic Young's modulus under each condition and weighting the prior distribution, the marginal probability of the dynamic Young's modulus is obtained.
[0033] (4) Calculation of posterior probability: According to the calculated prior probability, conditional probability and evidence factor, the posterior probability can be updated by inputting the data set D1. The update algorithm is as follows:
[0034] in, To import the dataset The calculated posterior probability is is the prior probability, For the dataset The calculated conditional probability is is the evidential factor; (5) Introduce the second data set D2 and calculate the new posterior probability: The posterior probability of the previous update As a priori, the calculation method is as follows:
[0035] in, To import the dataset The calculated posterior probability is is the prior probability, For the dataset The calculated conditional probability is is the evidential factor;
[0036] Here, we assume that the introduced data set D2 is the longitudinal wave velocity, and the observed longitudinal wave velocity state is medium, then the evidence factor The algorithm is as follows:
[0037]
[0038] In the above formula, D2 is the observed longitudinal wave velocity state, D1 is the observed dynamic Young's modulus state, is the conditional probability that the observed P-wave velocity state is between medium and rock hardness, is the posterior probability of rock hardness obtained during the first data update, By calculating the above formula, we can get the posterior probability of introducing the D2 data set for secondary update .
[0039] (6) Introduce the third data set D3 and calculate the new posterior probability: The posterior probability of the previous update As a priori, the calculation method is as follows:
[0040] in, To import the dataset The calculated posterior probability is is the prior probability, For the dataset The calculated conditional probability is is the evidential factor;
[0041] Here, we assume that the introduced data set D3 is density, and assume that the observed density state is high, then the evidence factor The algorithm is as follows:
[0042] In the above formula, D3 is the observed Poisson's ratio state, D2 is the observed longitudinal wave velocity state, D1 is the observed density state, is the conditional probability between the observed Poisson's ratio state being high and the rock hardness, is the posterior probability of rock hardness obtained during the second data update, By calculating the above formula, we can get the posterior probability of introducing the D3 dataset for three updates: .
[0043] At this point, a complete posterior probability update based on three data sets D1, D2, and D3 has been completed, and the generated posterior probability It represents the new probability distribution formed by the changes in concepts and probabilities after the prior probability is observed continuously for three data sets related to rock hardness.
[0044] Example 3 A rockburst disaster identification method based on geological inference process includes the following steps: Step S21, collecting unexcavated TSP data in front of the tunnel face and excavated geological data behind the tunnel face; The unexcavated TSP data in front of the tunnel face include longitudinal wave velocity, shear wave velocity, Poisson's ratio, density, static Young's modulus, dynamic Young's modulus and direct wave velocity; the excavated geological data behind the tunnel face include rock hardness, groundwater conditions, integrity and ground stress; Step S22, using the collected unexcavated TSP data in front of the tunnel face and the excavated geological data behind the tunnel face to infer the unexcavated geological data in front of the tunnel face using a geological parameter inference model; Inference models include:
[0045] in, : Posterior probability, the probability that the parameter θ is the true value after given the data D; : Likelihood function, which indicates the probability of data D appearing when parameter θ is true; : Prior probability refers to the original belief about parameter θ before observing the current data; : The marginal likelihood, also called the normalization constant or evidence factor, is the total probability of the observed data D occurring; Step S23, constructing a rockburst disaster identification model; The rockburst disaster identification model includes an input layer, a convolutional layer 1, a convolutional layer 2, a pooling layer, a Flatten layer, a fully connected layer, and an output layer connected in sequence; Step S24, pre-processing the unexcavated TSP data in front of the tunnel face and the inferred unexcavated geological data in front of the tunnel face to form a data set; The preprocessing includes the following steps: Step A1: normalize the unexcavated TSP data features in front of the tunnel face: The MinMaxScaler function is used to scale the numerical features in the unexcavated TSP data in front of the tunnel face to the interval [0,1]; The MinMaxScaler formula is:
[0046] in is the original data, and are the minimum and maximum values of the feature respectively; Step A2: Encode the inferred unexcavated geological data and disaster categories ahead of the tunnel face: The text features in the inferred unexcavated geological data ahead of the tunnel face are converted into numerical values by mapping; Step A3: Map target labels: Map the string "rockburst level" to a numeric label; Step A4: Split training set and test set: The dataset is divided into training set and test set in a ratio of 80:20; Step A5, one-hot encoding: Apply OneHotEncoder to the target variable, i.e., rockburst level, to convert the category label into a one-hot vector; Step S25: Model training, optimization and evaluation: The model is trained and optimized using an iterative training method, which can be specifically divided into initial model training and iterative training; Iterative evaluation includes 100-250 training iterations, using 10-20 samples per iteration, and splitting 20-30% of the data from the training data for validation to evaluate the performance of the model during training; Use the test set for evaluation, calculate the classification accuracy, predict and output the true category and predicted category of the test set, and select accuracy as the performance indicator; The specific steps of model training and optimization are: Step B1: Model compilation: The Adam optimizer performs initial configuration on the model, allowing the model to optimize parameters using an adaptive learning rate. The loss function is categorical_crossentropy, the goal is to minimize the error between the category output probability and the actual label; Step B2, callback function: Use ModelCheckpoint and EarlyStopping to help adjust the model, automatically save the best model, prevent overfitting and improve training efficiency; Step B3: Training model: The training set is input into the model for parameter optimization, and the performance is verified through multiple iterations of evaluation; The specific steps of model evaluation are: Step C1: Visualization of model training results: Use Matplotlib to plot the loss and accuracy curves during the training and validation phases; Step C2, model evaluation: Evaluate on the test set, calculate the classification accuracy, predict and output the true category and predicted category of the test set, and select accuracy as the performance indicator; Step S26, using the trained and optimized model to identify rockburst disasters; like Figure 4 As shown in the figure, after 150 iterations of training the rockburst disaster identification model, the loss rate of the model is roughly stable at around 0.2, while the accuracy of the model's validation set is around 92%, and the accuracy of the training set is around 88%, and the model fit is good.
[0047] After completing the statistics of the prior probability and the conditional probability, the present invention uses the method of using the prior data as the prior data of the next step. Each time the observed D1, D2, and D3 states are input, the prior probability can be combined with the influence of the three physical states to generate the posterior probability. The generated posterior probability is then used as the prior probability. By inputting the observed D1, D2, and D3 states multiple times, the prior probability can be updated multiple times, thereby realizing a dynamic update process of the prior probability combined with the influence of the observed state. According to the posterior probability distribution after each update, the state with the highest probability in the rock hardness state can be selected as the inference result. The geological data of the excavated tunnel is processed as a priori, that is, a general judgment of the rear of the tunnel. Then, TSP advanced geophysical exploration is used to collect TSP data behind the face, and the collected TSP data is input into the inference model as observation data. The input observation data will affect the priori according to the conditional probability relationship between TSP data and geological data to obtain the posterior, and the posterior is the geological data inference result of the unexcavated section corresponding to the input TSP data; then the updated posterior is used as the priori, and the subsequent TSP data is continuously input, so as to continuously update the posterior to obtain the geological data inference result, and realize the inference of the four indicators of rock hardness, groundwater conditions, integrity and ground stress of the geological data of the unexcavated part; The rockburst hazard identification model is used in combination with the collected excavated geological data behind the tunnel face and the inferred unexcavated geological data in front of the tunnel face to realize the rockburst hazard level prediction operation of the unexcavated part in front of the tunnel face, providing data support for tunnel geological construction.
[0048] The above-mentioned embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and substitutions made by technicians in this field on the basis of the present invention shall fall within the scope of protection required by the present invention.
Claims
1. A tunnel disaster identification method based on geological inference process, characterized in that: The following steps are involved: Step S1, collecting unexcavated TSP data in front of the tunnel face and excavated geological data behind the tunnel face; Step S2, using the collected unexcavated TSP data in front of the tunnel face and the excavated geological data behind the tunnel face to infer the unexcavated geological data in front of the tunnel face using a geological parameter inference model; Step S3, constructing a rockburst disaster identification model; Step S4, pre-processing the unexcavated TSP data in front of the tunnel face and the inferred unexcavated geological data in front of the tunnel face to form a data set; Step S5: Model training, optimization and evaluation: The model is trained and optimized using an iterative training method, which can be specifically divided into initial model training and iterative training; Use the test set for evaluation, calculate the classification accuracy, predict and output the true category and predicted category of the test set, and select accuracy as the performance indicator; Step S6: Using the trained and optimized model to identify rockburst disasters.
2. A tunnel disaster identification method based on geological inference process as claimed in claim 1, characterized in that: The unexcavated TSP data in front of the tunnel face in step S1 include longitudinal wave velocity, shear wave velocity, Poisson's ratio, density, static Young's modulus, dynamic Young's modulus and direct wave velocity; the excavated geological data behind the tunnel face include rock hardness, groundwater conditions, integrity and ground stress.
3. A tunnel disaster identification method based on geological inference process as claimed in claim 1, characterized in that: The Bayesian inference model in step S2 includes: ; in, : posterior probability, the probability that the parameter θ is the true value after given the data D; : Likelihood function, which indicates the probability of data D appearing when parameter θ is true; : Prior probability refers to the original belief about parameter θ before observing the current data; : The marginal likelihood, also called the normalizing constant or evidence factor, is the overall probability of the observation D occurring.
4. A tunnel disaster identification method based on geological inference process as claimed in claim 1, characterized in that: The rockburst disaster identification model in step S3 includes an input layer, a convolutional layer 1, a convolutional layer 2, a pooling layer, a Flatten layer, a fully connected layer and an output layer which are connected in sequence.
5. A tunnel disaster identification method based on geological inference process as claimed in claim 1, characterized in that: The pre-processing in step S4 comprises the following steps: Step A1: normalize the unexcavated TSP data features in front of the tunnel face: The MinMaxScaler function is used to scale the numerical features in the unexcavated TSP data in front of the tunnel face to the interval [0,1]; Step A2: Encode the inferred unexcavated geological data and disaster categories ahead of the tunnel face: The text features in the inferred unexcavated geological data ahead of the tunnel face are converted into numerical values by mapping; Step A3: Map target labels: Map the string "rockburst level" to a numerical label; Step A4: Split training set and test set: The dataset is divided into training set and test set in a ratio of 80:20; Step A5, one-hot encoding: OneHotEncoder is applied to the target variable, i.e., rockburst grade, to convert the category labels into one-hot vectors.
6. A tunnel disaster identification method based on geological inference process as claimed in claim 1, characterized in that: The model training and optimization in step S5 includes the following steps: Step B1: Model compilation: The Adam optimizer performs initial configuration on the model, allowing the model to optimize parameters using an adaptive learning rate. The loss function is categorical_crossentropy, the goal is to minimize the error between the category output probability and the actual label; Step B2, callback function: Use ModelCheckpoint and EarlyStopping to help adjust the model, automatically save the best model, prevent overfitting and improve training efficiency; Step B3: Training model: The training set is input into the model for parameter optimization, and the performance is evaluated and verified after multiple iterations.
7. A tunnel disaster identification method based on geological inference process as claimed in claim 1, characterized in that: The model evaluation in step S5 includes the following steps: Step C1: Visualization of model training results: Use Matplotlib to plot the loss and accuracy curves during the training and validation phases; Step C2, model evaluation: Evaluate on the test set, calculate the classification accuracy, predict and output the true category and predicted category of the test set, and select accuracy as the performance indicator.
8. A tunnel disaster identification method based on geological inference process as claimed in claim 6, characterized in that: The iterative evaluation includes 100-250 training iterations, using 10-20 samples in each iteration, and splitting 20-30% of the data from the training data for validation to evaluate the performance of the model during the training process.
9. A tunnel disaster identification method based on geological inference process as claimed in claim 5, characterized in that: The unexcavated TSP data features in front of the tunnel face are normalized using the MinMaxScaler function, where the MinMaxScaler formula is: ; in, is the original data, and are the minimum and maximum values of the feature, respectively.
Citation Information
Patent Citations
Underground engineering pathological block advanced prediction method based on TSP system and block theory
CN107832505A
Dynamic risk obtaining method for tunnel large deformation disasters
CN110210121A
Tunnel gushing risk grade fuzzy prediction method based on comprehensive advanced geological prediction technology
CN110705795A
Geological advanced fine forecasting method based on convolutional neural network multi-geophysical-exploration coupling
CN112130216A
Semi-supervised shield tunnel face geological type prediction method and system
CN113431635A
