Rockburst intensity forecasting method based on model-independent element learning improved RNN-DNN hybrid neural network

Abstract

The invention provides a rockburst intensity forecasting method based on a model-independent element learning improved RNN-DNN hybrid neural network, belongs to the technical field of rock-soil rockburst, and aims to solve the problems of low forecasting accuracy and insufficient cross-engineering scene adaptability in a traditional forecasting method. The method comprises the following steps: widely collecting multi-source rockburst influence factor data, determining a rockburst grade division standard, and constructing a rockburst database; carrying out missing value complementation through Bayesian interpolation to obtain first-level data; performing abnormal value elimination by using an ECOD abnormal detection algorithm to obtain secondary data; dividing secondary data into a training set and a verification set; adopting a model independent element learning framework to optimize initial parameters of the RNN-DNN hybrid neural network; and loading the initial parameters to the hybrid neural network, and carrying out model training by combining the training set and the verification set to obtain a rockburst intensity prediction model. According to the method, the accuracy and cross-scene universality of rockburst intensity prediction can be improved, and technical support is provided for disaster prevention and control of deep underground engineering.

Description

Technical Field

[0001] The present invention relates to the field of geotechnical engineering, and more particularly to a rockburst intensity prediction method based on a RNN-DNN hybrid neural network improved by model-independent meta-learning (MAML). Background Art

[0002] Rockbursts are common in various fields, including underground engineering, mining, and tunnel construction. Their sudden and destructive nature poses a serious challenge to the safety and economic viability of projects. Rockbursts are caused by the instantaneous release of energy in underground rocks due to changes in various environmental factors, such as stress, temperature, moisture content, and construction methods. These phenomena are often accompanied by violent manifestations such as vibration, sound waves, and rock fractures. This phenomenon not only threatens the safety of workers but can also damage equipment and delay construction schedules. Therefore, in-depth research on the triggering mechanisms and influencing factors of rockbursts is crucial for effectively predicting and controlling rockburst risks.

[0003] Traditional rockburst intensity prediction methods often rely on single monitoring indicators or empirical models, which suffer from the following technical drawbacks: First, rockburst causes are complex, involving heterogeneous data from multiple sources, including geological structure, geostress fields, and surrounding rock properties. Existing methods struggle to effectively integrate this heterogeneous engineering information, limiting prediction accuracy. Second, data in actual projects often suffer from missing values ​​and outliers, making traditional preprocessing methods inadequately robust and prone to noise. Third, traditional machine learning models have poor generalization capabilities in small sample sizes and multi-task scenarios, making them difficult to adapt to the dynamic demands of diverse engineering environments. These issues limit the accuracy and versatility of rockburst predictions. Summary of the Invention

[0004] In response to the above problems, the present 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 problems of limited accuracy and versatility of rockburst prediction in current engineering practice.

[0005] In order to achieve the above technical features, the purpose of the present invention is to achieve the following: a rockburst intensity prediction method based on model-independent meta-learning to improve the RNN-DNN hybrid neural network, the method comprising the following steps: Step 1: Collect rockburst influencing factors from different engineering sources, determine rockburst classification standards, and establish a database; Step 2: Use Bayesian interpolation to fill in the missing values ​​in the database and obtain primary data; Step 3: Use the ECOD algorithm to check the outliers in the primary data described in step 2, and remove the outliers to obtain the secondary data; Step 4: Divide the secondary data described in step 3 into a training set and a validation set; Step 5: Establish RNN-DNN neural network; Step 6: Input the training set and validation set described in step 4 into the model-agnostic meta-learning (MAML) algorithm to calculate the initial parameters of the RNN-DNN hybrid neural network; 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 a trained neural network; Step 8: Use the trained neural network described in step 7 to predict the rockburst data to obtain the rockburst grade.

[0006] Preferably, the rockburst influencing factor data in step 1 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 ; It also includes factors such as groundwater conditions, construction impacts, and rock properties; The rockburst grade classification standard is determined according to different situations.

[0007] Preferably, in step 2, Bayesian interpolation is used to fill in missing values ​​in the database, comprising the following steps: (1) Initialize missing values ​​and fill in missing values ​​to form a complete data set; (2) Establishing the Bayesian regression model Assume that the missing data obey the linear regression model as follows: ; Where: X obs For the observed data, X mis For missing data, β is the regression coefficient, which needs to be estimated. ε is the error term, which follows a zero-mean normal distribution; (3) Calculate the posterior distribution of the regression model parameters. In the Bayesian framework, given the observation data X obs and the prior distribution P( β ), use Bayes’ theorem to update the posterior distribution of the regression coefficients as shown below: ; Where: P ( X obs ∣ β) is the likelihood function, that is, given the regression coefficient β and observational data X obs The probability of P ( β ) is the prior distribution of the regression coefficient; (4) Sampling from the posterior distribution to obtain a set of regression coefficient samples, and using these regression coefficient samples to calculate the possible value of each missing value to generate multiple interpolation results; (5) Update the regression coefficients and the posterior distribution of missing values ​​in multiple iterations, using the interpolated values ​​from the previous step as input in each iteration until the change in the interpolated values ​​becomes stable or the predetermined maximum number of iterations is reached.

[0008] Preferably, the step 3 of using the ECOD algorithm to check the outliers in the database and remove the outliers includes the following steps: (1) Calculate the empirical cumulative distribution of each dimension of the data using the following formula: ; Where: F ( x ) is in x The empirical cumulative distribution value at ; I ( x ) is an indicator function, which is 1 if the condition in the brackets is true, otherwise it is 0; (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; (3) Determine whether a data point is an outlier based on the outlier score and remove the outlier.

[0009] Preferably, in step 4, the ratio of dividing the database into a training set and a validation set is 80% and 20%.

[0010] Preferably, the RNN-DNN neural network established in step 5 is a hybrid neural network using a recurrent layer in the first half and a fully connected layer in the second half.

[0011] Preferably, in step six, the training set and the validation set are input into the MAML algorithm, and the step of calculating the initial parameters of the RNN-DNN hybrid neural network also includes: the training set is used in the model parameter optimization process of the MAML algorithm, and the initial parameters of the RNN-DNN hybrid neural network are iteratively updated by the gradient descent method; the validation set is used as the basis for determining the early stopping mechanism of the MAML algorithm. When the loss function value on the validation set does not show a significant decrease after several consecutive iterations, the training process is terminated to prevent the occurrence of overfitting.

[0012] Preferably, the initial parameters in step six are the initial weights and initial biases of the neural network model.

[0013] The present invention has the following beneficial effects: 1. Efficient integration and cleaning of multi-source data: Using Bayesian interpolation and ECOD anomaly detection algorithms, we systematically address the common issues of missing values ​​and abnormal noise in rockburst data. Bayesian interpolation dynamically infers missing values ​​based on a probabilistic model, preserving the data distribution characteristics. The ECOD algorithm uses nonparametric statistics to identify multi-dimensional abnormal samples, avoiding the subjectivity of traditional thresholding methods and significantly improving data quality and model input reliability.

[0014] 2. Rapid model adaptation in small sample scenarios: The Model-Agnostic Meta-Learning (MAML) framework was introduced to optimize the initial parameters of the RNN-DNN hybrid neural network using task data from multiple engineering projects, enabling the model to rapidly adapt across scenarios. Compared to traditional training methods based on single-engineering data, MAML uses gradient aggregation during the meta-training phase, enabling the model to achieve high-precision predictions on new projects with only minimal fine-tuning, breaking through the bottleneck of traditional methods' reliance on small sample sizes.

[0015] 3. Collaborative modeling of time series and nonlinear features: The RNN module captures the temporal evolution of rockburst precursors, while the DNN module analyzes the nonlinear dynamics of multi-factor coupling. The combination of the two avoids the drawback of static models that ignore temporal information while enhancing the ability to deeply characterize complex rockburst mechanisms.

[0016] 4. Improve engineering practice efficiency: This method is simple and efficient, making it particularly suitable for multi-factor analysis in complex rock mass conditions. Compared with traditional methods, its results are more reliable and provide a more scientific basis for decision-making in engineering design.

[0017] 5. Strong adaptability: The method of the present invention can predict rockburst intensity and is faster and more versatile than traditional methods in terms of accuracy and applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention will be further described below with reference to the accompanying drawings and examples.

[0019] Figure 1 Flowchart of the rockburst intensity prediction method based on improved RNN-DNN hybrid neural network based on model-agnostic meta-learning (MAML).

[0020] Figure 2This is the specific flowchart of the MAML part in the rockburst intensity prediction method of improving the RNN-DNN hybrid neural network based on model-independent meta-learning (MAML).

[0021] Figure 3 This is a comparison of the model prediction result indicators in the embodiment and the prediction results of ordinary neural networks under the same data basis conditions. DETAILED DESCRIPTION

[0022] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] Example 1: An embodiment of the present invention provides a rockburst intensity prediction method based on an improved RNN-DNN hybrid neural network using model-independent meta-learning (MAML), the method comprising: Step 1: Collect rockburst influencing factors from various engineering sources, determine rockburst classification standards, and establish a database; Among them, the data of rock burst influencing factors include but are not limited to 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 , and can also collect influencing factors such as groundwater conditions, construction impact, rock properties, etc.; the rock burst grade classification standards can be determined according to different situations.

[0025] Step 2: Use Bayesian interpolation to fill in the missing values ​​in the database and obtain primary data; The specific steps include: (1) Initialize missing values ​​and fill in missing values ​​to form a complete data set; (2) Establishing the Bayesian regression model Assume that the missing data obey the linear regression model as follows: ; Where: X obs For the observed data, X mis For missing data, βis the regression coefficient, which needs to be estimated. ε is the error term, which follows a zero-mean normal distribution; (3) Calculate the posterior distribution of the regression model parameters. In the Bayesian framework, given the observation data X obs and the prior distribution P( β ), use Bayes’ theorem to update the posterior distribution of the regression coefficients as shown below: ; Where: P ( X obs ∣ β ) is the likelihood function, that is, given the regression coefficient β and observational data X obs The probability of P ( β ) is the prior distribution of the regression coefficient; (4) Sampling from the posterior distribution to obtain a set of regression coefficient samples, and using these regression coefficient samples to calculate the possible value of each missing value to generate multiple interpolation results; (5) Update the regression coefficients and the posterior distribution of missing values ​​in multiple iterations, using the interpolated values ​​from the previous step as input in each iteration until the change in the interpolated values ​​becomes stable or the predetermined maximum number of iterations is reached.

[0026] Step 3: Use the ECOD algorithm to check the outliers in the primary data described in step 2 and remove them to obtain secondary data; The specific steps include: (1) Calculate the empirical cumulative distribution of each dimension of the data using the following formula: ; Where: F ( x ) is in x The empirical cumulative distribution value at ; I ( x ) is an indicator function, which is 1 if the condition in the brackets is true, otherwise it is 0; (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; (3) Determine whether a data point is an outlier based on the outlier score and remove the outlier.

[0027] Step 4: Divide the secondary data described 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%.

[0028] Step 5: Establish RNN-DNN neural network; The RNN-DNN neural network established in step 5 is a hybrid neural network using a recurrent layer in the first half and a fully connected layer in the second half.

[0029] Step 6: Input the training set and validation set described in step 4 into the MAML algorithm to calculate the initial parameters of the RNN-DNN hybrid neural network; The training set is used in the model parameter optimization process of the MAML algorithm, and the initial parameters of the RNN-DNN hybrid neural network are iteratively updated through the gradient descent method; the validation set is used 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 after several consecutive iterations, the training process is terminated to prevent overfitting.

[0030] 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 a trained neural network; The initial parameters are the initial weights and initial biases of the neural network model.

[0031] Step 8: Use the trained neural network described in step 7 to predict the rockburst data to obtain the rockburst grade.

[0032] Example 2: The present invention is further described below with reference to specific embodiments, which specifically include the following steps: (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 The rock burst level is divided into levels 0 to 3, corresponding to no rock burst, slight rock burst, moderate rock burst and severe rock burst respectively. The relevant classification standards are basic common sense in the field of rock burst. Different classification standards can be adopted according to different projects. I will not go into details here. The relevant data are shown in Table 1 below: Table 1 Partial rockburst data of a certain project

[0033] (2) Bayesian interpolation is performed on the data in the database to eliminate the impact of missing data, as shown in Table 2.

[0034] Table 2 Partial interpolation data

[0035] (3) Calculate the empirical cumulative distribution of the data samples under each influencing factor (dimension), and calculate the abnormal value score of each data point according to the empirical cumulative distribution. The score calculation method is the negative logarithm of the empirical cumulative distribution function of the point.

[0036] (4) The outlier scores of each data sample in the six dimensions are added together to obtain the final score. In this embodiment, the outlier contamination ratio is 10%. The top 10% of samples are eliminated based on the outlier scores and contamination ratios to obtain the final database.

[0037] (5) The database is divided into a training set and a validation set with a ratio of 80% and 20%.

[0038] (6) Construct an RNN-DNN neural network. The neural network constructed in this embodiment specifically includes: the first three layers are recurrent layers, and the number of recurrent units from the first layer to the third layer is 96, 96, and 80 respectively; the last three layers are fully connected layers, and the number of neurons is 32, 16, and 16; the input layer dimension is 6, and the output layer dimension is 4; the optimizer selects Adam, and the loss function selects the cross entropy function.

[0039] (7) The training set and validation set are input into the MAML algorithm to calculate the initial parameters of the RNN-DNN hybrid neural network. The training set is used for 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 rounds, the MAML training is stopped, and the MAML training result before the 5 training rounds is taken as the initial parameters of the RNN-DNN hybrid neural network.

[0040] (8) Input the initial parameters into the RNN-DNN hybrid neural network and perform supervised learning optimization based on the training set and the validation set. After the training is completed, save the network weight parameters and build a rockburst intensity prediction model. The model can receive engineering geological parameter input and output rockburst grade classification results. In this embodiment, the performance indicators of the prediction model are as follows: Figure 3 The figure compares the performance gap between the rockburst intensity prediction method based on the improved RNN-DNN hybrid neural network using model-agnostic meta-learning (MAML) and the conventional method without optimizing the neural network under the same data conditions.

[0041] The above detailed description of the specific embodiment of the present invention is intended to be only one embodiment, and the present invention is not limited to the specific embodiment 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, any equivalent changes and modifications made without departing from the spirit and scope of the present invention should be included within the scope of the present invention.

Claims

1. A rockburst intensity prediction method based on model-independent meta-learning to improve RNN-DNN hybrid neural network, characterized by: The method comprises the following steps: Step 1: Collect rockburst influencing factors from different engineering sources, determine rockburst classification standards, and establish a database; Step 2: Use Bayesian interpolation to fill in the missing values ​​in the database and obtain primary data; Step 3: Use the ECOD algorithm to check the outliers in the primary data described in step 2, and remove the outliers to obtain the secondary data; Step 4: Divide the secondary data described in step 3 into a training set and a validation set; Step 5: Establish RNN-DNN neural network; Step 6: Input the training set and validation set described in step 4 into the model-agnostic meta-learning (MAML) algorithm to calculate the initial parameters of the RNN-DNN hybrid neural network; 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 a trained neural network; Step 8: Use the trained neural network described in step 7 to predict the rockburst data to obtain the rockburst grade.

2. The rockburst intensity prediction method based on model-independent meta-learning improved RNN-DNN hybrid neural network according to claim 1 is characterized in that: The rockburst influencing factor data in step 1 include 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 ; It also includes factors such as groundwater conditions, construction impacts, and rock properties; The rockburst grade classification standard is determined according to different situations.

3. The rockburst intensity prediction method based on model-independent meta-learning to improve RNN-DNN hybrid neural network according to claim 1 is characterized in that: The second step uses Bayesian interpolation to fill in the missing values ​​in the database, including the following steps: (1) Initialize missing values ​​and fill in missing values ​​to form a complete data set; (2) Establishing the Bayesian regression model Assume that the missing data obey the linear regression model as follows: ； Where: X obs For the observed data, X mis For missing data, β is the regression coefficient, which needs to be estimated. ε is the error term, which follows a zero-mean normal distribution; (3) Calculate the posterior distribution of the regression model parameters. In the Bayesian framework, given the observation data X obs and the prior distribution P( β ), use Bayes’ theorem to update the posterior distribution of the regression coefficients as shown below: ； Where: P ( X obs ∣ β ) is the likelihood function, that is, given the regression coefficient β and observational data X obs The probability of P ( β ) is the prior distribution of the regression coefficient; (4) Sampling from the posterior distribution to obtain a set of regression coefficient samples, and using these regression coefficient samples to calculate the possible value of each missing value to generate multiple interpolation results; (5) Update the regression coefficients and the posterior distribution of missing values ​​in multiple iterations, using the interpolated values ​​from the previous step as input in each iteration until the change in the interpolated values ​​becomes stable or the predetermined maximum number of iterations is reached.

4. The rockburst intensity prediction method based on model-independent meta-learning to improve RNN-DNN hybrid neural network according to claim 1 is characterized in that: The step 3 of using the ECOD algorithm to check for outliers in the database and remove outliers includes the following steps: (1) Calculate the empirical cumulative distribution of each dimension of the data using the following formula: ； Where: F ( x ) is in x The empirical cumulative distribution value at ; I ( x ) is an indicator function, which is 1 if the condition in the brackets is true, otherwise it is 0; (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; (3) Determine whether a data point is an outlier based on the outlier score and remove the outlier.

5. The rockburst intensity prediction method based on model-independent meta-learning to improve RNN-DNN hybrid neural network according to claim 1 is characterized in that: In step 4, the database is divided into a training set and a validation set in a ratio of 80% and 20%.

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

7. The rockburst intensity prediction method based on model-independent meta-learning to improve RNN-DNN hybrid neural network according to claim 1 is characterized in that: In step six, the training set and the validation set are input into the MAML algorithm, and the step of calculating the initial parameters of the RNN-DNN hybrid neural network also includes: the training set is used in the model parameter optimization process of the MAML algorithm, and the initial parameters of the RNN-DNN hybrid neural network are iteratively updated by the gradient descent method; the validation set is used as the basis for determining the early stopping mechanism of the MAML algorithm. When the loss function value on the validation set does not show a significant decrease after several consecutive iterations, 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 RNN-DNN hybrid neural network according to claim 1 is characterized in that: The initial parameters in step six are the initial weights and initial biases of the neural network model.

Images