A Rockburst Intensity Prediction Method Based on Model-Independent Meta-Learning Improved RNN-DNN Hybrid Neural Network

Abstract

This invention provides a rockburst intensity prediction method based on a model-independent meta-learning (MIL) improved RNN-DNN hybrid neural network, belonging to the field of rockburst technology. It aims to solve the problems of low prediction accuracy and insufficient adaptability across engineering scenarios in traditional prediction methods. The method includes the following steps: extensively collecting multi-source rockburst influencing factor data and determining rockburst level classification standards to construct a rockburst database; completing missing values ​​using Bayesian imputation to obtain primary data; using the ECOD anomaly detection algorithm to remove outliers to obtain secondary data; dividing the secondary data into training and validation sets; optimizing the initial parameters of the RNN-DNN hybrid neural network using a model-independent meta-learning framework; loading the initial parameters into the hybrid neural network, and training the model using the training and validation sets to obtain a rockburst intensity prediction model. This invention can improve the accuracy and cross-scenario applicability of rockburst intensity prediction, providing technical support for disaster prevention and control in deep underground engineering.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering, and more specifically to a method for predicting rockburst intensity based on an improved RNN-DNN hybrid neural network using Model Independent Meta-Learning (MAML). Background Technology

[0002] Rockbursts are prevalent in various fields such as underground engineering, mining, and tunnel construction. Their suddenness and destructiveness pose a severe challenge to the safety and economy of engineering projects. A rockburst is an instantaneous energy release phenomenon that occurs in underground rock under varying environmental factors such as stress, temperature, water content, and construction methods. It is often accompanied by violent manifestations such as vibration, sound waves, and rock mass fracturing. This phenomenon not only affects the safety of workers but may also lead to equipment damage and construction delays. Therefore, in-depth research into the inducing mechanism and influencing factors of rockbursts is crucial for effectively predicting and controlling rockburst risks.

[0003] Traditional methods for predicting rockburst intensity often rely on single monitoring indicators or empirical models, resulting in the following technical limitations: First, rockburst causes are complex, involving multi-source heterogeneous data such as geological structures, geostress fields, and surrounding rock properties. Existing methods struggle to effectively integrate this heterogeneous engineering information, limiting prediction accuracy. Second, real-world engineering data often contains missing values ​​and outliers, making traditional preprocessing methods insufficiently robust and prone to introducing noise interference. Third, traditional machine learning models exhibit poor generalization ability in small-sample, multi-task scenarios, making it difficult to adapt to the dynamic changes in different engineering environments. These issues restrict the accuracy and versatility of rockburst prediction. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a rockburst intensity prediction method based on an improved RNN-DNN hybrid neural network using model-independent meta-learning, aiming to solve the limitations in the accuracy and versatility of rockburst prediction in current engineering practice.

[0005] To achieve the above-mentioned technical features, the objective of this invention is as follows: a rockburst intensity prediction method based on model-independent meta-learning improved RNN-DNN hybrid neural network, the method comprising the following steps:

[0006] Step 1: Collect rockburst influencing factors from different engineering sources, determine the rockburst classification standards, and establish a database;

[0007] Step 2: Use Bayesian imputation to complete the missing values ​​in the database and obtain the first-level data;

[0008] Step 3: Use the ECOD algorithm to check for outliers in the primary data described in Step 2, and remove the outliers to obtain the secondary data;

[0009] Step 4: Divide the secondary data mentioned in Step 3 into a training set and a validation set;

[0010] Step 5: Build the RNN-DNN neural network;

[0011] Step Six: Input the training set and validation set described in Step Four into the Model Independent Meta-Learning (MAML) algorithm to calculate the initial parameters of the RNN-DNN hybrid neural network;

[0012] Step 7: Input the initial parameters described in Step 6 into the RNN-DNN hybrid neural network, and train the neural network using the training set and validation set described in Step 4 to obtain the trained neural network;

[0013] Step 8: Using the trained neural network described in Step 7, predict the rockburst data to be predicted and obtain the rockburst level.

[0014] Preferably, the rockburst influencing factor data in step one includes the maximum tangential stress of the surrounding rock. MTS Uniaxial compressive strength of rock UCS uniaxial tensile strength of rock UTS Rock elastic strain energy index WET Rock stress coefficient SCF and rock brittleness coefficient B ;

[0015] This also includes groundwater conditions, construction impacts, and factors affecting rock properties;

[0016] The criteria for classifying rockburst levels are determined independently based on different circumstances.

[0017] Preferably, step two uses Bayesian imputation to complete missing values ​​in the database, including the following steps:

[0018] (1) Initialize missing values, fill in the missing values ​​first to form a complete dataset;

[0019] (2) Establish a Bayesian regression model assuming the missing data follows a linear regression model as follows:

[0020] ;

[0021] In the formula: X obs For the observed data, X mis For missing data, β These are regression coefficients, which need to be estimated. ε It is the error term, which follows a zero-mean normal distribution;

[0022] (3) Calculate the posterior distribution of the regression model parameters. In the Bayesian framework, based on the given observation data... X obs The prior distribution P of the regression model β The posterior distribution of the regression coefficients is updated using Bayes' theorem, as shown in the following equation:

[0023] ;

[0024] In the formula: P ( X obs | β ) is the likelihood function, i.e., given the regression coefficients β and observation data X obs The probability, P ( β ) is the prior distribution of the regression coefficients;

[0025] (4) Sample from the posterior distribution to obtain a set of regression coefficient samples. Calculate the possible values ​​of each missing value using these regression coefficient samples to generate multiple imputation results.

[0026] (5) Iterate and update the regression coefficients and the posterior distribution of missing values ​​multiple times. Each iteration uses the imputed value from the previous step as input until the change of the imputed value tends to stabilize or the predetermined maximum number of iterations is reached.

[0027] Preferably, step three, which involves using the ECOD algorithm to check for outliers in the database and removing them, includes the following steps:

[0028] (1) Calculate the empirical cumulative distribution of each dimension of the data, using the following formula:

[0029] ;

[0030] In the formula: F ( x ) is in x The empirical cumulative distribution value at that location; I ( x The parentheses are an indicator function; if the condition inside the parentheses is true, the value is 1; otherwise, it is 0.

[0031] (2) Use the empirical distribution to estimate the tail probability of each data point in each dimension, and calculate the outlier score of each data point by aggregating the estimated tail probabilities across dimensions. The specific calculation method is to calculate the negative logarithm of the tail probability.

[0032] (3) Determine whether a data point is an outlier based on the outlier score and remove outliers.

[0033] Preferably, in step four, the database is divided into a training set and a validation set in a ratio of 80% and 20%, respectively.

[0034] Preferably, the RNN-DNN neural network established in step five is a hybrid neural network that uses recurrent layers in the first half and fully connected layers in the second half.

[0035] Preferably, step six, which involves inputting the training set and validation set into the MAML algorithm to calculate the initial parameters of the RNN-DNN hybrid neural network, further includes: using the training set for the model parameter optimization process of the MAML algorithm, iteratively updating the initial parameters of the RNN-DNN hybrid neural network using gradient descent; and using the validation set as the basis for the early stopping mechanism of the MAML algorithm, terminating the training process when the loss function value on the validation set does not show a significant decrease for several consecutive iterations to prevent overfitting.

[0036] Preferably, the initial parameters in step six are the initial weights and initial biases of the neural network model.

[0037] The present invention has the following beneficial effects:

[0038] 1. Efficient integration and cleaning of multi-source data:

[0039] By employing Bayesian imputation and the ECOD anomaly detection algorithm, this system systematically addresses the common problems of missing values ​​and anomalous noise in rockburst data. Bayesian imputation dynamically infers missing values ​​based on a probabilistic model, preserving the data distribution characteristics; the ECOD algorithm identifies multi-dimensional anomalous samples through nonparametric statistics, avoiding the subjectivity of traditional thresholding methods and significantly improving data quality and model input reliability.

[0040] 2. Fast model adaptation in scenarios with small sample sizes:

[0041] The Model-Independent Meta-Learning (MAML) framework is introduced to optimize the initial parameters of the RNN-DNN hybrid neural network using task data from multiple engineering sources, giving the model the ability to quickly adapt across different scenarios. Compared with the traditional single-engineering data training mode, MAML, through gradient aggregation in the meta-training stage, enables the model to achieve high-precision prediction with only a few fine-tunings when facing new engineering scenarios, breaking through the bottleneck of traditional methods' dependence on small sample data.

[0042] 3. Co-modeling of temporal and nonlinear characteristics:

[0043] The RNN module captures the temporal evolution of rockburst precursors, while the DNN module analyzes the nonlinear dynamic mechanism involving multiple coupled factors. The combination of these two modules avoids the limitation of static models that ignore the temporal dimension, while simultaneously enhancing the ability to deeply represent complex rockburst mechanisms.

[0044] 4. Improve engineering practice efficiency:

[0045] This method is simple and efficient in its calculation process, and is particularly suitable for multi-factor analysis under complex rock mass conditions. Compared with traditional methods, its results are more realistic and reliable, providing a more scientific basis for engineering design decisions.

[0046] 5. Highly adaptable:

[0047] The method of this invention can predict rockburst intensity and is faster and more versatile than traditional methods in terms of accuracy and applicability. Attached Figure Description

[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0049] Figure 1 This is a flowchart illustrating a method for predicting rockburst intensity based on an improved RNN-DNN hybrid neural network using Model-Independent Meta-Learning (MAML).

[0050] Figure 2 This is a flowchart illustrating the MAML component of an improved RNN-DNN hybrid neural network-based rockburst intensity prediction method based on Model-Independent Meta-Learning (MAML).

[0051] Figure 3 This example compares the model prediction results with those of a regular neural network under the same data conditions. Detailed Implementation

[0052] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example 1:

[0055] This invention provides a method for predicting rockburst intensity based on an improved RNN-DNN hybrid neural network using Model-Independent Meta-Learning (MAML). The method includes:

[0056] Step 1: Collect information on rockburst influencing factors from various engineering sources, determine rockburst classification standards, and establish a database;

[0057] Among them, the data on factors affecting rockburst include, but are not limited to, the maximum tangential stress of the surrounding rock. MTS Uniaxial compressive strength of rock UCSuniaxial tensile strength of rock UTS Rock elastic strain energy index WET Rock stress coefficient SCF and rock brittleness coefficient B It can also collect influencing factors such as groundwater conditions, construction impact, and rock properties; the criteria for determining the rockburst level can be determined independently according to different situations.

[0058] Step 2: Use Bayesian imputation to complete the missing values ​​in the database and obtain the first-level data;

[0059] Specifically, the steps include the following:

[0060] (1) Initialize missing values, fill in the missing values ​​first to form a complete dataset;

[0061] (2) Establish a Bayesian regression model assuming the missing data follows a linear regression model as follows:

[0062] ;

[0063] In the formula: X obs For the observed data, X mis For missing data, β These are regression coefficients, which need to be estimated. ε It is the error term, which follows a zero-mean normal distribution;

[0064] (3) Calculate the posterior distribution of the regression model parameters. In the Bayesian framework, based on the given observation data... X obs The prior distribution P of the regression model β The posterior distribution of the regression coefficients is updated using Bayes' theorem, as shown in the following equation:

[0065] ;

[0066] In the formula: P ( X obs | β ) is the likelihood function, i.e., given the regression coefficients β and observation data X obs The probability, P ( β ) is the prior distribution of the regression coefficients;

[0067] (4) Sample from the posterior distribution to obtain a set of regression coefficient samples. Calculate the possible values ​​of each missing value using these regression coefficient samples to generate multiple imputation results.

[0068] (5) Iterate and update the regression coefficients and the posterior distribution of missing values ​​multiple times. Each iteration uses the imputed value from the previous step as input until the change of the imputed value tends to stabilize or the predetermined maximum number of iterations is reached.

[0069] Step 3: Use the ECOD algorithm to check for outliers in the primary data described in Step 2 and remove them to obtain the secondary data;

[0070] Specifically, the steps include the following:

[0071] (1) Calculate the empirical cumulative distribution of each dimension of the data, using the following formula:

[0072] ;

[0073] In the formula: F ( x ) is in x The empirical cumulative distribution value at that location; I ( x The parentheses are an indicator function; if the condition inside the parentheses is true, the value is 1; otherwise, it is 0.

[0074] (2) Use the empirical distribution to estimate the tail probability of each data point in each dimension, and calculate the outlier score of each data point by aggregating the estimated tail probabilities across dimensions. The specific calculation method is to calculate the negative logarithm of the tail probability.

[0075] (3) Determine whether a data point is an outlier based on the outlier score and remove outliers.

[0076] Step 4: Divide the secondary data mentioned in Step 3 into a training set and a validation set; the ratio of the training set to the validation set is 80% and 20%, respectively.

[0077] Step 5: Build the RNN-DNN neural network;

[0078] The RNN-DNN neural network established in step five is a hybrid neural network that uses recurrent layers in the first half and fully connected layers in the second half.

[0079] Step Six: Input the training set and validation set described in Step Four into the MAML algorithm to calculate the initial parameters of the RNN-DNN hybrid neural network;

[0080] The training set is used for the model parameter optimization process of the MAML algorithm, and the initial parameters of the RNN-DNN hybrid neural network are iteratively updated using the gradient descent method. The validation set serves as the basis for the early stopping mechanism of the MAML algorithm. When the loss function value on the validation set does not show a significant decrease for several consecutive iterations, the training process is terminated to prevent overfitting.

[0081] Step 7: Input the initial parameters described in Step 6 into the RNN-DNN hybrid neural network, and train the neural network using the training set and validation set described in Step 4 to obtain the trained neural network;

[0082] The initial parameters are the initial weights and initial biases of the neural network model.

[0083] Step 8: Using the trained neural network described in Step 7, predict the rockburst data to be predicted and obtain the rockburst level.

[0084] Example 2:

[0085] The present invention will be further described below with reference to specific embodiments, which specifically include the following steps:

[0086] (1) In this embodiment, the maximum tangential stress of the surrounding rock is collected. MTS Uniaxial compressive strength of rock UCS uniaxial tensile strength of rock UTS Rock elastic strain energy index WET Rock stress coefficient SCF and rock brittleness coefficient B Rockburst levels are classified into 0 to 3, corresponding to no rockburst, minor rockburst, moderate rockburst, and severe rockburst, respectively. The relevant classification standards are basic common knowledge in the field of rockburst. Different classification standards can be adopted according to different projects, which will not be elaborated here. Some relevant data are shown in Table 1 below:

[0087] Table 1. Partial Rockburst Data for a Certain Project

[0088]

[0089] (2) Perform Bayesian imputation on the data in the database to eliminate the impact of missing data, as shown in Table 2.

[0090] Table 2 Partially interpolated data

[0091]

[0092] (3) Calculate the empirical cumulative distribution of the data sample under each influencing factor (dimension), and calculate the outlier score of each data point according to the empirical cumulative distribution. The score is calculated as the negative logarithm of the empirical cumulative distribution function of that point.

[0093] (4) Add up the scores of outliers in the six dimensions for each data sample to get the final score. In this embodiment, the outlier contamination ratio is 10%. The top 10% of the samples are removed based on the outlier scores and contamination ratio to obtain the final database.

[0094] (5) Divide the database into training set and validation set, with a ratio of 80% and 20%.

[0095] (6) Construct an RNN-DNN neural network. In this embodiment, the neural network specifically includes: the first three layers are recurrent layers with the number of recurrent units from the first layer to the third layer being 96, 96, and 80 respectively; the last three layers are fully connected layers with the number of neurons being 32, 16, and 16 respectively; the input layer dimension is 6 and the output layer dimension is 4; the optimizer is Adam and the loss function is cross-entropy function.

[0096] (7) Input the training set and validation set into the MAML algorithm to calculate the initial parameters of the RNN-DNN hybrid neural network. The training set is used for the meta-learning training of MAML, and the validation set is used to calculate the validation loss. In this embodiment, if the validation loss does not decrease in 5 training sessions, the MAML training is stopped, and the MAML training result before 5 training sessions is taken as the initial parameters of the RNN-DNN hybrid neural network.

[0097] (8) The initial parameters are input into an RNN-DNN hybrid neural network, and supervised learning optimization is performed based on the training and validation sets. After training, the network weight parameters are saved to construct a rockburst intensity prediction model. This model can receive engineering geological parameter inputs and output rockburst level classification results. In this embodiment, the performance indicators of the prediction model are as follows: Figure 3 As shown in the figure, the performance difference between the rockburst intensity prediction method based on model-independent meta-learning (MAML) and the ordinary method without optimizing the neural network under the same data conditions is compared.

[0098] The specific embodiments of the present invention have been described in detail above, but these are merely one example, and the present invention is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, all equivalent changes and modifications made without departing from the spirit and scope of the present invention should be covered within the scope of the present invention.

Claims

1. A rock burst intensity prediction method based on model-independent meta-learning to improve an RNN-DNN hybrid neural network, characterized in that, The method comprises the following steps: Step one: collect rock burst influencing factors from different engineering sources, determine rock burst grade classification standards, and establish a database; Step two: use Bayesian interpolation to complete the missing values in the database to obtain primary data; Step three: use the ECOD algorithm to check the abnormal values in the primary data in step two, and remove the abnormal values to obtain secondary data; Step four: divide the secondary data in step three into a training set and a validation set; Step five: establish an RNN-DNN neural network; Step six: input the training set and the validation set in step four into the model-agnostic meta-learning (MAML) algorithm to calculate the initial parameters of the RNN-DNN hybrid neural network; Step seven: input the initial parameters in step six into the RNN-DNN hybrid neural network, train the neural network using the training set and the validation set in step four, and obtain the trained neural network; Step eight: use the trained neural network in step seven to predict the rock burst data to be predicted and obtain the rock burst grade.

2. The rockburst intensity prediction method based on model-independent meta-learning to improve the RNN-DNN hybrid neural network according to claim 1, characterized in that, The rock burst influencing factor data in the step one includes maximum tangential stress of surrounding rock MTS , uniaxial compressive strength of rock UCS , uniaxial tensile strength of rock UTS , elastic strain energy index of rock WET , stress coefficient of rock SCF and brittleness coefficient of rock B ; It also includes groundwater conditions, construction impact conditions and rock property influencing factors; The determination of the rock burst grade classification standards is self-determined according to different conditions.

3. The rockburst intensity prediction method based on model-independent meta-learning to improve the RNN-DNN hybrid neural network according to claim 1, characterized in that, The use of Bayesian interpolation to complete the missing values in the database in step two comprises the following steps: (1) initialize the missing values, first fill in the missing values to form a complete data set; (2) establish a Bayesian regression model assuming that the missing data obeys a linear regression model as follows: ； where: X obs is observed data, X mis is missing data, β is a regression coefficient that needs to be estimated, (4) sample from the posterior distribution to obtain a sample of regression coefficients, calculate the possible values of each missing value through these samples of regression coefficients, and generate multiple interpolation results; is an error term that is normally distributed with zero mean. (3) Compute the posterior distribution of the regression model parameters, in a Bayesian framework, given the observed data X obs and the prior distribution P( β ) of the regression model, the posterior distribution of the regression coefficients is updated using Bayes' theorem, as follows: ； where: P ( X obs ∣ β ) is the likelihood function, i.e. the probability of the observed data β given the regression coefficients X obs , (p P ( β ) is the prior distribution of the regression coefficients; (5) update the regression coefficients and the posterior distribution of the missing values multiple times, and use the interpolation value of the last step as the input each time until the interpolation value tends to be stable or reaches a predetermined maximum number of iterations. The use of the ECOD algorithm to check the abnormal values in the database in step three comprises the following steps:

4. The rockburst intensity prediction method based on model-independent meta-learning to improve the RNN-DNN hybrid neural network according to claim 1, characterized in that, (1) calculate the empirical cumulative distribution of each dimension of the data, and the calculation method is as follows: (2) use the empirical distribution to estimate the tail probability of each data point in each dimension, and calculate the abnormal value score of each data point by aggregating the estimated tail probability across dimensions, and the specific calculation method is to calculate the negative logarithm of the tail probability; ； wherein: F ( x ) is the empirical cumulative distribution value at x ; I ( x ) is an indicator function, which is 1 if the condition within the parentheses is true, and 0 otherwise. (3) determine whether the data point is an abnormal value according to the abnormal value score, and remove the abnormal value. The proportion of the database divided into the training set and the validation set in step four is 80% and 20%.

5. The rockburst intensity prediction method based on model-independent meta-learning to improve the RNN-DNN hybrid neural network according to claim 1, characterized in that, The RNN-DNN neural network established in step five is a hybrid neural network using a recurrent layer in the first half and a fully connected layer in the second half.

6. The rockburst intensity prediction method based on model-independent meta-learning to improve the RNN-DNN hybrid neural network according to claim 1, characterized in that, ​ 7. The rockburst intensity prediction method based on model-agnostic meta-learning to improve the RNN-DNN hybrid neural network according to claim 1, characterized in that, The step six inputs the training set and the validation set into the MAML algorithm, and the step of calculating the initial parameters of the RNN-DNN hybrid neural network further includes: the training set is used for the model parameter optimization process of the MAML algorithm, the initial parameters of the RNN-DNN hybrid neural network are iteratively updated by the gradient descent method, and the validation set is used as a judgment basis of an early stopping mechanism of the MAML algorithm; when the loss function value on the validation set does not appear significant decline for several times of continuous iteration, the training process is terminated to prevent the occurrence of overfitting.

8. The rockburst intensity prediction method based on model-independent meta-learning to improve the RNN-DNN hybrid neural network according to claim 1, characterized in that, The initial parameters in the step six are initial weights and initial biases of the neural network model.

Images