A method and system for early warning of rockburst based on seismic-acoustic-electrical fusion using automated machine learning

Abstract

This invention discloses a method and system for early warning of rockbursts based on automatic machine learning and the fusion of seismic, acoustic, and electrical signals. The method includes: collecting multi-source monitoring data such as microseismic events, acoustic emission, and electromagnetic radiation from a field-deployed microseismic and acoustic-electrical monitoring system to provide data support for model construction; proposing two modules based on an automatic machine learning framework: automatic seismic, acoustic, and electrical modeling and prediction, and multi-parameter integrated early warning; the MAEP module automatically identifies key temporal features in the multi-source monitoring data, constructs a high-performance prediction model integrating multiple algorithms, accurately identifies the nonlinear and co-evolutionary characteristics of multi-source signals, and realizes the prediction of high-energy rockburst events; the MAEW module constructs an early warning index based on the prediction results, adopts an ensemble learning algorithm, comprehensively reflects the abrupt change characteristics and abnormal enhancement trends of seismic, acoustic, and electrical signals, and can automatically optimize the model structure and parameters according to data changes to achieve efficient, intelligent, and real-time early warning without human intervention.

Description

Technical Field

[0001] This invention relates to the field of rockburst early warning, and in particular to a rockburst early warning method and system based on automatic machine learning using seismic-acoustic-electrical fusion. Background Technology

[0002] As mineral resources are increasingly mined at greater depths, deep mines and tunnel engineering face more complex geological structures, high ground stress, high ground temperature, and multi-field coupling effects, posing significant engineering challenges. Against this backdrop, rockbursts, as sudden dynamic disasters caused by high-energy accumulation, exhibit severe destructive forms and complex triggering mechanisms. They are often accompanied by high-energy microseismic events, acoustic emission from rock fractures, and abnormal electromagnetic wave releases, posing a serious threat to the underground working environment. Statistical data shows that the frequency and intensity of coal mine rockburst events have increased significantly in recent years. Therefore, constructing a scientific, efficient, and stable rockburst prediction and early warning technology system has become a key technological bottleneck in ensuring safe production in deep mines.

[0003] During the formation and occurrence of rockbursts, coal and rock masses undergo an evolutionary process from deformation and fracturing to energy release, accompanied by various precursory features such as increased microseismic frequency, enhanced acoustic emission energy, and abnormal electromagnetic radiation. To achieve efficient identification of these abnormal signals, coal mines commonly construct multi-source monitoring systems based on microseismic, acoustic emission, and electromagnetic radiation. These systems possess advantages such as non-invasiveness and strong real-time performance, and can comprehensively reflect the energy evolution process of coal and rock masses, serving as an important source of information for rockburst early warning. Although traditional methods such as the empirical threshold method and the three-parameter comprehensive method have achieved disaster early warning to some extent, it is difficult to extract universally effective early warning features due to differences in physical mechanisms, inconsistent sampling, and feature coupling among the seismic, acoustic, and electrical data. In recent years, intelligent technologies such as deep learning have demonstrated significant advantages in modeling complex data in rockburst prediction, effectively mining nonlinear, multi-scale disaster evolution characteristics. Patent CN119670963A proposes a method for constructing a large-scale model for predicting rockbursts based on multimodal data. This method constructs a multimodal precursor pattern sequence, combines it with Transformer to predict the probability of rockburst levels, and integrates geological and mining information to achieve highly adaptable and accurate risk assessment. Patent CN119644435A discloses a rockburst prediction method and system driven by source inversion data. Based on source data inversion and the Informer prediction model, it integrates risk assessment to achieve multidimensional intelligent prediction and disaster mechanism analysis of the time, location, magnitude, and evolution process of rockbursts. However, existing research often focuses on modeling single signal sources, failing to achieve seismic-acoustic-electrical data fusion, and relies on experts manually designing network structures and feature variables, resulting in high computational costs, complex deployment, and difficulty in meeting real-time monitoring and early warning needs. The emergence of AutoML technology provides a new approach to solving this problem, automatically completing feature engineering, model selection, hyperparameter optimization, and integrated training, improving development efficiency and generalization ability. As a high-performance AutoML framework launched by Amazon, AutoGluon possesses powerful multi-model fusion capabilities, automatic parameter tuning mechanisms, and end-to-end modeling capabilities, making it suitable for multi-source, high-dimensional data modeling tasks. Therefore, constructing a seismic-acoustic-electrical multi-source fusion method for predicting rockbursts based on AutoGluon can effectively fuse seismic, acoustic, and electrical data and automatically and quickly select the optimal prediction model, thereby improving the accuracy and real-time performance of rockburst prediction. Summary of the Invention

[0004] This solution addresses the problems and needs raised above by proposing a method and system for early warning of rockbursts based on automatic machine learning and fusion of seismic, acoustic, and electrical signals. The method achieves the aforementioned technical objectives and brings about several other technical benefits by adopting the following technical features.

[0005] One objective of this invention is to propose a seismic-acoustic-electrical fusion method for early warning of rockbursts based on automated machine learning, comprising the following steps:

[0006] S10: Based on the microseismic system and acoustic-electric system deployed on site, construct a multi-source monitoring network covering the working face, and screen and acquire high-quality microseismic data, acoustic emission data and electromagnetic radiation data;

[0007] S20: Perform preprocessing operations on the microseismic data, acoustic emission data and electromagnetic radiation data collected on site to form a seismic-acoustic-electromagnetic fusion prediction dataset with a unified timestamp and input structure, and divide the prediction dataset into training set, validation set and test set;

[0008] S30: Construct an automatic modeling and prediction module for seismic, acoustic, and electrical signals based on the AutoGluon automatic machine learning platform. Using preprocessed multi-source signals as input, the module automatically searches and optimizes various prediction model structures and hyperparameter configurations using the AutoGluon framework. It constructs a prediction system sensitive to nonlinear and multi-scale dynamic changes, uses the training set to fit and learn the model, evaluates the model performance and tunes the hyperparameters using the validation set, and selects the model with the best MASE index on the validation set within a set time. The model is then applied to the test set to achieve accurate prediction of seismic, acoustic, and electrical signals.

[0009] S40: After normalizing the microseismic data, acoustic emission data and electromagnetic radiation data collected on site, the key feature parameters of electromagnetic radiation, acoustic emission and microseismic signals are extracted according to time units. Combined with the artificially constructed early warning index, a comprehensive earthquake, acoustic and electromagnetic early warning dataset is constructed and divided into training set, validation set and test set.

[0010] S50: Construct a multi-parameter integrated early warning module based on AutoGluon, using key features of multi-source monitoring data as input. Integrate regression algorithms to automatically complete feature processing, model selection, and hyperparameter optimization. Construct a high-precision regression prediction model for rockburst early warning index. Use the training set to learn and fit various models. Automatically search and optimize hyperparameter combinations during training. Iteratively test and evaluate performance on the validation set to ensure the optimal model structure and parameter configuration. Input the test set into the regression prediction model and evaluate the model performance using the RMSE index to achieve automatic selection and accurate prediction of the optimal model.

[0011] In addition, the seismic-acoustic-electrical fusion rockburst early warning method and system based on automatic machine learning according to the present invention may also have the following technical features:

[0012] In one example of the present invention, in step S20, the seismic-acoustic-electrical data collected on site are preprocessed to construct and divide a seismic-acoustic-electrical fusion prediction dataset, including the following steps:

[0013] S21: Missing and outlier values ​​in acoustic emission and electromagnetic radiation data are imputed using the moving average, time step is standardized, and the data is normalized; microseismic data retains energy greater than 1.0 × 10⁻⁶. 3 For event J, missing values ​​are filled with the daily average and a uniform step size; finally, the three types of data are merged into a multi-source input dataset with a unified structure.

[0014] S22: 88% of the multi-source input dataset is allocated as the training set and 12% as the test set. During the training phase, automatic prediction technology is used, and the training and validation subsets are divided based on K-fold cross-validation for model training, evaluation, and hyperparameter optimization to improve generalization ability.

[0015] In one example of the present invention, in step S30, constructing a seismoacoustic-electrical automatic modeling and prediction module based on the automated machine learning AutoGluon includes the following steps:

[0016] S31: Data Import: Construct an automatic modeling and prediction module for seismic, acoustic and electrical signals based on AutoGluon automatic machine learning. Each time series of preprocessed multi-source signals is input into the system as an independent modeling item. The prediction range is flexibly set according to specific task requirements to ensure that the model input meets the prediction requirements of different levels.

[0017] S32: Model Training: The MAEP module is based on the AutoGluon automated machine learning framework and uses an automated model management mechanism for modeling. Within the set time budget, the system explores multiple model structures and hyperparameter configurations to the maximum extent. For each input time series, multiple prediction models are first fitted and learned using the training set, and the model performance is ranked based on the MASE error index on the validation set. The system automatically selects the best-performing model for prediction. After the model training is completed, all models and their results are saved in the specified folder.

[0018] S33: Model Analysis and Evaluation: The MAEP module uses image processing to visually compare the true and predicted values, providing an intuitive qualitative assessment of the prediction trend and correlation. At the same time, it uses the mean absolute scaling error (MASE) for quantitative evaluation. The smaller the MASE value, the better the prediction performance. This is used to compare the merits of different models and select the best model. The test set is then input into the optimal model to achieve accurate prediction and early warning of future seismic and acoustic signals.

[0019] In one example of the present invention, in step S33, it is assumed that y represents the true value. Let n represent the predicted value and n be the number of samples. Then, the formula for calculating MASE is:

[0020]

[0021] In one example of the present invention, step S40 specifically includes the following steps:

[0022] S41: The collected microseismic data, acoustic emission data, and electromagnetic radiation data are normalized and preprocessed. Key characteristic parameters of electromagnetic radiation, acoustic emission, and microseismic signals are extracted using "days" as the time unit, including the maximum average electromagnetic radiation intensity, the average electromagnetic radiation intensity, the number of electromagnetic radiation events, the maximum average acoustic emission intensity, the average acoustic emission intensity, the number of acoustic emission events, the maximum energy value of microseismic events, and the average energy value of microseismic events. Combined with on-site monitoring records and expert experience scoring, a unified early warning index is calculated.

[0023] S42: A single time-series sample is constructed by combining the daily seismic, acoustic, and electrical monitoring features with the corresponding early warning index to form a comprehensive seismic, acoustic, and electrical early warning dataset. 80% of the samples are divided into training sets and validated by K-fold cross-validation. The remaining 20% ​​of the samples, after removing the early warning index, are used as the test set.

[0024] In one example of the present invention, in step S50, constructing a comprehensive early warning module for seismic, acoustic, and electrical multi-parameters based on AutoGluon automatic machine learning includes the following steps:

[0025] S51: Data Import: Construct a multi-parameter integrated early warning module for seismic, acoustic, and electrical signals based on the AutoGluon automatic machine learning model, and use preprocessed multi-source monitoring data as input. This data covers key characteristic parameters of signals such as electromagnetic radiation, acoustic emission, and microseismic activity, and uses the early warning index as the prediction target. It is uniformly constructed into a data format for structured regression tasks and input into the system for subsequent modeling and analysis.

[0026] S52: Model Training: The MAEW module is based on the AutoGluon automatic machine learning framework for modeling. This module intelligently generates and trains multiple prediction models based on the values ​​of multiple feature columns in the data table to predict the value of the target column. MAEW integrates multiple types of algorithms through training set fitting learning to fully explore the nonlinear and high-dimensional feature correlations in the data. Secondly, it evaluates and ranks the performance of each model based on the root mean square error index on the validation set, and integrates multiple high-performing sub-models through a multi-model integration strategy, and finally automatically constructs a prediction model with optimal generalization ability.

[0027] S53: Model Analysis and Evaluation: During the prediction phase, the MAEW module automatically applies the best-performing model combination to perform regression prediction on the test set; during the evaluation process, the root mean square error (RMSE) is selected as the core performance indicator, which reflects the accuracy of the model by measuring the overall deviation between the predicted value and the true value; the smaller the RMSE, the better the prediction effect.

[0028] In one example of the present invention, in step S53, during the model analysis and evaluation process, the negative RMSE value is used as the scoring criterion to rank the models based on their performance, ensuring that the final output model has optimal performance; wherein, it is assumed that y represents the true value. Let n represent the predicted value and n be the number of samples. Then, the formula for calculating RMSE is:

[0029]

[0030] Another objective of this invention is to propose a seismic-acoustic-electrical fusion-based early warning system for rockbursts based on automated machine learning, comprising:

[0031] The seismic, acoustic, and electromagnetic data acquisition module is configured to build a multi-source monitoring network covering the working face based on the microseismic system and acoustic and electrical system deployed on site, and to screen and acquire high-quality microseismic data, acoustic emission data, and electromagnetic radiation data.

[0032] The seismic-acoustic-electromagnetic prediction dataset construction module is configured to preprocess the microseismic data, acoustic emission data and electromagnetic radiation data collected on site to form a seismic-acoustic-electromagnetic fusion prediction dataset with a unified timestamp and input structure, and divide the prediction dataset into training set, validation set and test set.

[0033] The automatic modeling and prediction module for seismic, acoustic, and electrical signals is configured to build an automatic modeling and prediction module for seismic, acoustic, and electrical signals based on the AutoGluon automatic machine learning framework. It takes preprocessed multi-source signals as input, and relies on the AutoGluon framework to realize the automatic search and optimization of various prediction model structures and hyperparameter configurations. It builds a prediction system that is sensitive to nonlinear and multi-scale dynamic changes, uses the training set to fit and learn the model, and then uses the validation set to evaluate the model performance and tune the hyperparameters. Within a set time, it selects the model with the best MASE index on the validation set and applies it to the test set to achieve accurate prediction of seismic, acoustic, and electrical signals.

[0034] The earthquake, acoustic, and electromagnetic early warning dataset construction module is configured to normalize the microseismic data, acoustic emission data, and electromagnetic radiation data collected on-site, extract the key feature parameters of electromagnetic radiation, acoustic emission, and microseismic signals according to time units, combine them with artificially constructed early warning indices, construct an earthquake, acoustic, and electromagnetic comprehensive early warning dataset, and divide it into training set, validation set, and test set.

[0035] The integrated early warning module for seismic, acoustic, and electrical data is configured to build an AutoGluon-based integrated early warning module for seismic, acoustic, and electrical data. It takes key features of multi-source monitoring data as input, integrates regression algorithms, and automatically completes feature processing, model selection, and hyperparameter optimization to build a high-precision regression prediction model for rockburst early warning index. It uses the training set to learn and fit various models, automatically searches and optimizes hyperparameter combinations during training, and performs iterative testing and performance evaluation on the validation set to ensure that the optimal model structure and parameter configuration are obtained. The test set is input into the regression prediction model, and the model performance is evaluated using the RMSE index, thus achieving automatic selection of the optimal model and accurate prediction.

[0036] In one example of the present invention, the seismic-acoustic-electric prediction dataset construction module includes:

[0037] The data preprocessing unit is configured to perform moving average imputation on missing and outlier values ​​in acoustic emission and electromagnetic radiation data, standardize the time step, and retain energy values ​​greater than 1.0 × 10⁻⁶ in microseismic data. 3 For event J, missing values ​​are filled with the daily average and a uniform step size; finally, the three types of data are merged into a multi-source input dataset with a unified structure.

[0038] The dataset partitioning unit is configured to divide 88% of the multi-source input dataset into a training set and 12% into a test set. During the training phase, automatic prediction technology is used, and the training and validation subsets are divided based on K-fold cross-validation for model training, evaluation, and hyperparameter optimization to improve generalization ability.

[0039] In one example of the present invention, the automatic modeling and prediction module for seismic acoustics and electricity includes:

[0040] The data import unit is configured to build an automatic modeling and prediction module for seismic acoustics and electricity based on the automatic machine learning AutoGluon. Each time series of the preprocessed multi-source signal is input into the system as an independent modeling item, and the prediction range can be flexibly set according to the specific task requirements to ensure that the model input meets the prediction requirements of different levels.

[0041] The model training unit, configured as the MAEP module, is based on the AutoGluon automated machine learning framework and employs an automated model management mechanism for modeling. Within the set time budget, the system explores various model structures and hyperparameter configurations to the maximum extent. For each input time series, the system trains multiple models and ranks their performance based on the MASE error metric on the validation set, automatically selecting the best-performing model for prediction. After model training is complete, all models and their results are saved in a designated folder.

[0042] The model analysis and evaluation unit, configured as the MAEP module, uses image methods to visually compare the true and predicted values, providing an intuitive qualitative assessment of the prediction trend and correlation. At the same time, it uses the mean absolute scaling error (MASE) for quantitative evaluation. The smaller the MASE value, the better the prediction performance, which is used to compare the merits of different models and select the best model.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. This invention uses multi-source data fusion technology of seismic, acoustic and electrical sources for rockburst prediction and early warning, which significantly improves the accuracy of multi-source signal fusion modeling, effectively alleviates the potential impact of differences between heterogeneous multi-source signals on the consistency of monitoring results, optimizes the integration process of information from different signal sources, and thus improves the accuracy and efficiency of rockburst prediction.

[0045] 2. This invention innovatively introduces an automated machine learning framework, enabling the automated construction and integration of multiple models, significantly reducing reliance on manual intervention and simplifying the modeling process. Through automated feature engineering, hyperparameter optimization, and model selection, it not only improves modeling efficiency but also enhances the generalization ability and accuracy of the prediction model, thereby improving overall prediction performance.

[0046] 3. This invention significantly outperforms traditional machine learning methods in terms of prediction accuracy, computational efficiency, and model stability, and can accurately identify high-risk periods for rockbursts. Through optimized modeling processes and data processing strategies, it demonstrates excellent generalization ability and stability, possessing broad application prospects, especially in prediction tasks in complex and dynamic environments where it exhibits strong adaptability.

[0047] The preferred embodiments of the invention will be described in more detail below with reference to the accompanying drawings, so as to facilitate an understanding of the features and advantages of the invention. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0049] Figure 1 This is a flowchart of a seismic-acoustic-electrical fusion rockburst early warning method based on automatic machine learning according to an embodiment of the present invention;

[0050] Figure 2 This is a diagram illustrating a specific deployment scheme for the seismic-acoustic-electrical fusion-based early warning method for rockbursts based on automatic machine learning, according to an embodiment of the present invention.

[0051] Figure 3The flowchart of the Automatic Seismo-Acoustic-Electrical Modeling and Prediction (MAEP) module according to an embodiment of the present invention is shown below.

[0052] Figure 4 This is a flowchart of the Multi-Parameter Integrated Early Warning (MAEW) module according to an embodiment of the present invention;

[0053] Figure 5 This is a graph showing the raw data and prediction results of microseismic data, acoustic emission data, and electromagnetic radiation data according to an embodiment of the present invention;

[0054] Figure 6 This is a diagram showing the automatic early warning results of the MAEW module early warning index according to an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0056] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “comprising” or “including” mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as “connected” or “linked” are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships; these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0057] According to a first aspect of the present invention, a seismo-acoustic-electrical fusion method for early warning of rockbursts is provided, such as... Figure 1 , Figure 2 As shown, it includes the following steps:

[0058] S10: Based on the SOS microseismic system and KJ796 acoustic-electric system deployed on site, a multi-source monitoring network covering the working face is constructed. By setting thresholds and sampling parameters, multi-channel signals are automatically acquired and remotely transmitted. High-quality microseismic data, acoustic emission data and electromagnetic radiation data are screened and obtained to provide data support for fusion modeling and rockburst early warning.

[0059] S20: Perform preprocessing operations such as cleaning, standardization and time alignment on the microseismic data, acoustic emission data and electromagnetic radiation data (MS, AE and EMR) collected on site to form a seismic-acoustic-electric fusion prediction dataset with a unified timestamp and input structure. The prediction dataset is then divided into training set, validation set and test set for automatic modeling and prediction.

[0060] S30: Construct a Seismoacoustic-Electrical Automatic Modeling and Prediction (MAEP) module based on AutoGluon automatic machine learning. Using preprocessed multi-source signals as input, the module automatically searches and optimizes various prediction model structures and hyperparameter configurations based on the AutoGluon framework. It constructs a prediction system sensitive to nonlinear and multi-scale dynamic changes, uses the training set to fit and learn the model, and then evaluates the model performance and tunes the hyperparameters through the validation set. Within a set time, it selects the model with the best MASE index on the validation set and applies it to the test set to achieve accurate prediction of seismoacoustic-electric signals.

[0061] S40: After normalizing the microseismic data, acoustic emission data and electromagnetic radiation data collected on site, the key feature parameters of electromagnetic radiation, acoustic emission and microseismic signals are extracted according to the time unit of "day". Combined with the artificially constructed early warning index, a comprehensive earthquake, acoustic and electromagnetic early warning dataset is constructed and divided into training set, validation set and test set.

[0062] S50: Construct a multi-parameter integrated early warning (MAEW) module based on AutoGluon. Taking key features of multi-source monitoring data as input, it integrates regression algorithms such as XGBoost and random forest to automatically complete feature processing, model selection, and hyperparameter optimization. It constructs a high-precision regression prediction model for rockburst early warning index, uses the training set to learn and fit various models, automatically searches and optimizes hyperparameter combinations during training, and performs iterative testing and performance evaluation on the validation set to ensure that the optimal model structure and parameter configuration are obtained. The test set is input into the regression prediction model, and the model performance is evaluated using the RMSE index, realizing the automatic selection of the optimal model and accurate prediction.

[0063] This early warning method uses multi-source data fusion technology (seismic, acoustic, and electrical) to predict and warn of rockbursts. It significantly improves the accuracy of multi-source signal fusion modeling, effectively mitigates the potential impact of differences between heterogeneous multi-source signals on the consistency of monitoring results, and optimizes the integration process of information from different signal sources, thereby improving the accuracy and efficiency of rockburst prediction.

[0064] This early warning method innovatively introduces an automated machine learning framework, enabling the automated construction and integration of multiple models. This significantly reduces reliance on manual intervention and simplifies the modeling process. Through automated feature engineering, hyperparameter optimization, and model selection, it not only improves modeling efficiency but also enhances the generalization ability and accuracy of the prediction model, thereby improving overall prediction performance.

[0065] This early warning method significantly outperforms traditional machine learning methods in terms of prediction accuracy, computational efficiency, and model stability, and can accurately identify high-risk periods for rockbursts. Through optimized modeling processes and data processing strategies, it demonstrates excellent generalization ability and stability, and has broad application prospects, especially in prediction tasks in complex and dynamic environments where it exhibits strong adaptability.

[0066] In one example of the present invention, in step S20, the seismic-acoustic-electrical data collected on site are preprocessed to construct and divide a seismic-acoustic-electrical fusion prediction dataset, including the following steps:

[0067] S21: Missing and outlier values ​​in acoustic emission and electromagnetic radiation data were imputed using a 5-minute moving average, and a standardized 6-hour time step was implemented; microseismic (MS) data with energy greater than 1.0 × 10⁻⁶ was retained. 3 For event J, missing values ​​are filled with the daily average and a uniform step size; finally, the three types of data are merged into a multi-source input dataset with a unified structure.

[0068] S22: After preprocessing the seismic, acoustic, and electrical data, 88% of the multi-source input dataset is divided into a training set and 12% into a test set. During the training phase, automatic prediction technology is used, and the training and validation subsets are divided based on K-fold cross-validation for model training, evaluation, and hyperparameter optimization to improve generalization ability.

[0069] In one example of the present invention, in step S30, as Figure 3 As shown, the construction of the Seismoacoustic-Electrical Automatic Modeling and Prediction (MAEP) module based on AutoGluon automatic machine learning includes the following steps:

[0070] S31: Data Import: Construct a Seismic-Acoustic-Electrical Automatic Modeling and Prediction (MAEP) module based on AutoGluon automatic machine learning. Input each time series of the preprocessed multi-source signal as an independent modeling item into the system, and flexibly set the prediction range according to specific task requirements (e.g., predict the data for the next 20 days with a sampling interval of 6 hours) to ensure that the model input meets the prediction requirements of different levels.

[0071] S32: Model Training: The MAEP module is based on the AutoGluon automated machine learning framework and uses an automated model management mechanism for modeling. Within the set time budget, the system explores various model structures and hyperparameter configurations, including deep neural networks and regression models, to the maximum extent. For each input time series, multiple prediction models are first fitted and learned using the training set, and the model performance is ranked based on the MASE error index on the validation set. The system automatically selects the best-performing model for prediction. After model training is completed, all models and their results are saved in a specified folder. Each model supports flexible setting of the prediction interval length and outputs the predicted mean and predicted values ​​at different quantiles, providing a 90% confidence interval to meet the dual needs of short-term anomaly identification and medium- to long-term trend monitoring.

[0072] S33: Model Analysis and Evaluation: The MAEP module uses image processing to visually compare the true and predicted values, providing an intuitive qualitative assessment of the prediction trend and correlation. At the same time, it uses the mean absolute scaling error (MASE) for quantitative evaluation. The smaller the MASE value, the better the prediction performance. It can be used to compare the merits of different models and select the best model. The test set is then input into the optimal model to achieve accurate prediction and early warning of future seismic and acoustic signals.

[0073] In one example of the present invention, in step S33, it is assumed that y represents the true value. Let n represent the predicted value and n be the number of samples. Then, the formula for calculating MASE is:

[0074]

[0075] In one example of the present invention, step S40 specifically includes the following steps:

[0076] S41: The collected microseismic data, acoustic emission data, and electromagnetic radiation data are normalized and preprocessed. Key characteristic parameters of electromagnetic radiation, acoustic emission, and microseismic signals are extracted using "day" as the time unit, including maximum average electromagnetic radiation intensity (EMR MA), average electromagnetic radiation intensity (EMR AA), number of electromagnetic radiation events (EMR events), maximum average acoustic emission intensity (AE MA), average acoustic emission intensity (AE AA), number of acoustic emission events (AE events), maximum microseismic energy value (MS MP), and average energy value of microseismic events (MS AP). Combined with field monitoring records and expert experience scoring, a unified early warning index is calculated.

[0077] S42: A single time-series sample is constructed by combining the daily seismic, acoustic, and electrical monitoring features with the corresponding early warning index to form a comprehensive seismic, acoustic, and electrical early warning dataset. 80% of the samples are divided into training sets and validated by K-fold cross-validation. The remaining 20% ​​of the samples, after removing the early warning index, are used as the test set.

[0078] In one example of the present invention, in step S50, a multi-parameter integrated early warning (MAEW) module based on automated machine learning AutoGluon is constructed, such as... Figure 4 As shown, it includes the following steps:

[0079] S51: Data Import: Construct a multi-parameter integrated early warning (MAEW) module based on automatic machine learning AutoGluon, using preprocessed multi-source monitoring data as input; this data covers key characteristic parameters of signals such as electromagnetic radiation, acoustic emission and microseismic activity, and uses the early warning index as the prediction target, uniformly constructing it into a data format for structured regression tasks, and inputting it into the system for subsequent modeling and analysis;

[0080] S52: Model Training: The MAEW module is based on the AutoGluon automated machine learning framework for modeling. As an automated modeling method suitable for regression supervised learning tasks, MAEW can automatically complete the modeling process without manual feature engineering, hyperparameter tuning, and model selection. This module can intelligently generate and train multiple prediction models based on the values ​​of multiple feature columns in a data table to predict the value of the target column. MAEW integrates various types of algorithms (including neural networks, tree models, etc.) through training set fitting and learning, comprehensively mining the nonlinear and high-dimensional feature correlations in the data, thereby improving prediction performance. Secondly, it evaluates and ranks the performance of each model based on the root mean square error (RMSE) metric on the validation set, and fuses multiple high-performing sub-models through a multi-model ensemble strategy, ultimately automatically constructing a prediction model with optimal generalization ability.

[0081] S53: Model Analysis and Evaluation: During the prediction phase, the MAEW module automatically applies the best-performing model combination to perform regression prediction on the test set; during the evaluation process, the root mean square error (RMSE) is selected as the core performance indicator, which reflects the accuracy of the model by measuring the overall deviation between the predicted value and the true value; the smaller the RMSE, the better the prediction effect.

[0082] In one example of the present invention, in step S53, during the model analysis and evaluation process, the negative RMSE value is used as the scoring criterion to rank the models based on their performance, ensuring that the final output model has optimal performance; wherein, it is assumed that y represents the true value. Let n represent the predicted value and n be the number of samples. Then, the formula for calculating RMSE is:

[0083]

[0084] According to a second aspect of the present invention, a seismic-acoustic-electrical fusion rockburst early warning system based on automatic machine learning includes:

[0085] The seismic-acoustic-electric data acquisition module is configured to build a multi-source monitoring network covering the working face based on the SOS microseismic system and KJ796 acoustic-electric system deployed on site. By setting thresholds and sampling parameters, it realizes automatic acquisition and remote transmission of multi-channel signals, and filters and obtains high-quality microseismic data, acoustic emission data and electromagnetic radiation data, providing data support for fusion modeling and rockburst early warning.

[0086] The seismic-acoustic-electric prediction dataset construction module is configured to perform preprocessing operations such as cleaning, standardization, and time alignment on microseismic data, acoustic emission data, and electromagnetic radiation data (MS, AE, and EMR) collected on-site, forming a seismic-acoustic-electric fusion prediction dataset with a unified timestamp and input structure. The prediction dataset is then divided into training, validation, and test sets for automatic model building and prediction.

[0087] The automatic modeling and prediction module for seismoacoustic-electric signals is configured to build an automatic modeling and prediction (MAEP) module based on AutoGluon automatic machine learning. It takes preprocessed multi-source signals as input, and relies on the AutoGluon framework to realize the automatic search and optimization of various prediction model structures and hyperparameter configurations. It builds a prediction system that is sensitive to nonlinear and multi-scale dynamic changes, uses the training set to fit and learn the model, and then uses the validation set to evaluate the model performance and tune the hyperparameters. Within a set time, it selects the model with the best MASE index on the validation set and applies it to the test set to achieve accurate prediction of seismoacoustic-electric signals.

[0088] The earthquake, acoustic, and electromagnetic early warning dataset construction module is configured to extract key feature parameters of electromagnetic radiation, acoustic emission, and microseismic signals from field-collected microseismic data, acoustic emission data, and electromagnetic radiation data after normalization processing, using "days" as the time unit. Combined with artificially constructed early warning indices, it constructs a comprehensive earthquake, acoustic, and electromagnetic early warning dataset and divides it into training, validation, and test sets.

[0089] The multi-parameter integrated early warning module for seismic, acoustic, and electrical data is configured to build a multi-parameter integrated early warning (MAEW) module based on AutoGluon. It takes key features from multi-source monitoring data as input, integrates regression algorithms such as XGBoost and Random Forest, and automatically completes feature processing, model selection, and hyperparameter optimization. It constructs a high-precision regression prediction model for the rockburst early warning index, uses the training set to learn and fit various models, automatically searches and optimizes hyperparameter combinations during training, and performs iterative testing and performance evaluation on the validation set to ensure the optimal model structure and parameter configuration. The test set is input into the regression prediction model, and the RMSE index is used to evaluate the model performance, achieving automatic selection of the optimal model and accurate prediction.

[0090] This early warning system uses multi-source data fusion technology (seismic, acoustic, and electrical) to predict and warn of rockbursts. It significantly improves the accuracy of multi-source signal fusion modeling, effectively mitigates the potential impact of differences between heterogeneous multi-source signals on the consistency of monitoring results, and optimizes the integration process of information from different signal sources, thereby improving the accuracy and efficiency of rockburst prediction.

[0091] This early warning system innovatively introduces an automated machine learning framework, enabling the automated construction and integration of multiple models. This significantly reduces reliance on manual intervention and simplifies the modeling process. Through automated feature engineering, hyperparameter optimization, and model selection, it not only improves modeling efficiency but also enhances the generalization ability and accuracy of the prediction model, thereby improving overall prediction performance.

[0092] This early warning system significantly outperforms traditional machine learning methods in terms of prediction accuracy, computational efficiency, and model stability, and can accurately identify high-risk periods for rockbursts. Through optimized modeling processes and data processing strategies, it demonstrates excellent generalization ability and stability, possessing broad application prospects, especially in prediction tasks in complex and dynamic environments where it exhibits strong adaptability.

[0093] In one example of the present invention, the seismic-acoustic-electric prediction dataset construction module includes:

[0094] The data preprocessing unit is configured to impute missing and outlier values ​​in acoustic emission and electromagnetic radiation data using a 5-minute moving average, standardize the time step to a uniform 6-hour time step, and retain microseismic (MS) data with energy greater than 1.0 × 10⁻⁶. 3For event J, missing values ​​are filled with the daily average and a uniform step size; finally, the three types of data are merged into a multi-source input dataset with a unified structure.

[0095] The dataset partitioning unit is configured to divide 88% of the multi-source input dataset into a training set and 12% into a test set after the seismic, acoustic, and electrical data preprocessing. During the training phase, automatic prediction technology is used, and the training and validation subsets are divided based on K-fold cross-validation for model training, evaluation, and hyperparameter optimization to improve generalization ability.

[0096] In one example of the present invention, the automatic modeling and prediction module for seismic acoustics and electricity includes:

[0097] The data import unit is configured to build a Seismic-Acoustic-Electrical Automatic Modeling and Prediction (MAEP) module based on AutoGluon automatic machine learning. Each time series of the preprocessed multi-source signal is input into the system as an independent modeling item, and the prediction range can be flexibly set according to specific task requirements (such as predicting data for the next 20 days with a sampling interval of 6 hours) to ensure that the model input meets the prediction requirements of different levels.

[0098] The model training unit, configured as the MAEP module, is based on the AutoGluon automated machine learning framework and employs an automated model management mechanism for modeling. Within a set time budget, the system explores various model structures and hyperparameter configurations, including deep neural networks and regression models, to the maximum extent. For each input time series, the system trains multiple models and ranks their performance based on the MASE error metric on the validation set, automatically selecting the best-performing model for prediction. After model training is complete, all models and their results are saved in a designated folder. Each model supports flexible setting of the prediction interval length and outputs the predicted mean and predicted values ​​at different quantiles, providing a 90% confidence interval to meet the dual needs of short-term anomaly identification and medium- to long-term trend monitoring.

[0099] The model analysis and evaluation unit, configured as the MAEP module, uses image methods to visually compare the true and predicted values, providing an intuitive qualitative assessment of the prediction trend and correlation. At the same time, it uses the mean absolute scaling error (MASE) for quantitative evaluation. The smaller the MASE value, the better the prediction performance. It can be used to compare the merits of different models and select the best model.

[0100] Specific Cases

[0101] On-site collection environment:

[0102] Acoustic and electrical monitoring data were collected from field records at Hongyang No. 3 Mine from January to June 2019. The mine deployed acoustic and electrical monitoring equipment based on the KJ796 system to achieve real-time perception of underground safety risks. Acoustic emission sensors were installed on the roadway wall using anchor bolts, while electromagnetic radiation sensors were suspended from the roadway side and pointed towards the coal face using a non-contact method. The acoustic and electrical system automatically acquired, preprocessed, and extracted indicators based on a set triggering mechanism. The signals used for analysis included: acoustic emission amplitude, acoustic emission ring count, electromagnetic radiation amplitude, and electromagnetic radiation ring count. A total of 8 monitoring points were deployed at the mine working face, with the specific deployment plan as follows: Figure 2 As shown, the spacing between the sensors is dynamically adjusted according to the progress of the working face. #1001, #1005, #1009, and #1013 are electromagnetic radiation sensors; #1003, #1007, #1011, and #1015 are acoustic emission sensors; and #10, #11, #12, and #13 are microseismic sensors. The acoustic emission and electromagnetic radiation sensors are coupled to the acoustic-electrical data acquisition station through the KJ system substation, while the microseismic sensors are coupled to the microseismic data acquisition station through the SOS system substation. The acoustic-electrical data acquisition station and the microseismic data acquisition station are coupled to the terminal unit. The acoustic-electrical monitoring data and the microseismic monitoring data are analyzed on the terminal unit using acoustic-electrical monitoring analysis software and microseismic monitoring analysis software, respectively.

[0103] Microseismic monitoring data were also collected from field records at Hongyang No. 3 Mine from January to June 2019. The microseismic monitoring system was the SOS system developed by the Polish Mining Research Institute. Due to limitations in geological conditions, propagation paths, and noise interference, the signals received by different sensors showed significant differences, exhibiting strong inconsistency and complexity. To improve data quality, the original sensor records were first screened, selecting records with clear signals as valid samples, and using manually labeled microseismic events as a reference. After screening and calibration, monitoring data from sensors with IDs #10, #11, #12, and #13 at working face 1208 from January 8, 2019 to June 20, 2019 were selected for training, verification, and testing. During the monitoring period, working face 1208 recorded 648 large-energy microseismic events, with energy distributions ranging from 1.0 × 10⁻⁶. 4 J to 1.0×10 5 Between J, nine other high-energy events were recorded, all with energies exceeding 1.0 × 10⁻⁶. 5J. Each microseismic event was jointly acquired and recorded by four sensors. Microseismic events with energies greater than 1.0 × 10³ J were selected to construct a dataset containing 2628 valid microseismic signals, ensuring class balance in subsequent model training. The daily maximum microseismic energy and the daily cumulative number of microseismic events were also recorded. The relative distance between the electromagnetic radiation and acoustic emission sensors was maintained at approximately 0.5 m, with each sensor uploading a statistical value every 30 seconds on average. A final dataset containing 1,261,440 valid acoustic emission signals and 1,261,440 valid electromagnetic radiation signals was constructed, recording the daily maximum acoustic emission amplitude, number of acoustic emission rings, average acoustic emission amplitude, maximum acoustic emission energy, average acoustic emission energy, maximum electromagnetic radiation amplitude, number of electromagnetic radiation rings, average electromagnetic radiation amplitude, maximum electromagnetic radiation energy, and average electromagnetic radiation energy.

[0104] Analysis of prediction results:

[0105] like Figure 5 , Figure 6 As shown, the blue curve represents the historical true values ​​in the test set, and the green curve represents the future true values ​​in the test set. The red shaded area corresponds to the 90% confidence interval, indicating that the true value has a 90% probability of falling within this area, thus reflecting the reliability and stability of the prediction results.

[0106] Figure (a) shows that the MAEP technique performs well in microseismic (MS) data prediction, with most actual values ​​falling within the 90% confidence interval of the predicted results, indicating that the model effectively captures the long-term trend of the time series. Before the high-energy event on June 8th, the predicted results showed a steady upward trend, demonstrating its early warning capability. Although there were deviations on June 11th and 13th, the model still responded promptly to sudden changes, reflecting its sensitivity to drastic fluctuations, thanks to the model's ability to dynamically adjust parameters.

[0107] In Figure (b), MAEP's predictions of acoustic emission (AE) data also have high reliability, especially with high mean predictions before June 8, reflecting potential energy accumulation. The model adapts slightly slower after the event, but predictions become more accurate after June 15, demonstrating its ability to gradually adapt to new trends.

[0108] Figure (c) illustrates the predictive performance of MAEP on electromagnetic radiation (EMR) data, with most of the actual values ​​falling within the confidence interval. Before the event, the predicted mean is high and consistent with the actual value; after the event, it captures signal changes and provides corresponding adjustments, thus enhancing model reliability.

[0109] As can be seen from Figures (b) and (c), although AE and EMR did not increase synchronously, they both exhibited early warning signals such as "slow rise" or "maintaining a high level" before the event, reflecting the energy evolution process of the rock mass. Therefore, multi-source signal fusion helps to improve the accuracy of early warning.

[0110] Analysis of early warning results:

[0111] Depend on Figure 6 It is evident that the MAEW module generally fits the changing trend of the early warning index well, with the actual and predicted values ​​trending in unison and without systematic bias. However, at certain abrupt changes, such as May 22, 2019 (actual value 0.46, predicted value only 0.1), the model's response to abnormal fluctuations is insufficient, possibly due to a small number of abnormal samples or bias in expert scoring. In cases of extremely high values, such as June 7 and 8, 2019, the predicted results are lower than expected, indicating a weak generalization ability of the model for rare high values, which is related to the training dataset being concentrated in the 0.2–0.4 range. Nevertheless, MAEW still accurately captured the upward trend of the high-energy event on June 8, validating its early warning effectiveness.

[0112] Overall, the MAEW module performs stably in conventional volatility modeling and effectively predicts early warning index trends, demonstrating strong practicality. Although there is still room for improvement in responding to sudden anomalies and high-risk periods, its overall predictive performance is satisfactory.

[0113] The foregoing has described in detail, with reference to preferred embodiments, the exemplary implementation of the seismic-acoustic-electrical fusion rockburst early warning method and system based on automatic machine learning proposed in this invention. However, those skilled in the art will understand that, without departing from the concept of this invention, various modifications and alterations can be made to the above specific embodiments, and various combinations can be made to the various technical features and structures proposed in this invention, without exceeding the protection scope of this invention, which is determined by the appended claims.

Claims

1. A method for rock burst early warning based on automatic machine learning, characterized in that , comprising the following steps: S10: Based on the field layout of the microseismic system and the acoustic-electric system, a multi-source monitoring network covering the working face is constructed, and high-quality microseismic data, acoustic emission data and electromagnetic radiation data are obtained; S20: The microseismic data, acoustic emission data and electromagnetic radiation data collected on site are preprocessed to form a seismic acoustic electric fusion prediction data set with unified time stamp and input structure, and the prediction data set is divided into training set, validation set and test set; S30: Construct a seismic acoustic electric automatic modeling and prediction module based on automatic machine learning AutoGluon, input the preprocessed multi-source signal, rely on AutoGluon framework to realize automatic search and optimization of multiple prediction model structures and hyperparameter configuration, build a prediction system sensitive to nonlinear and multi-scale dynamic changes, use the training set to fit the model, then evaluate the model performance through the validation set and optimize the hyperparameters, select the model with the optimal MASE index on the validation set within a specified time, and apply it to the test set to realize accurate prediction of seismic acoustic electric signal; S40: After normalizing the microseismic data, acoustic emission data and electromagnetic radiation data collected on site, extract the key feature parameters of electromagnetic radiation, acoustic emission and microseismic signal according to time unit, combine the artificially constructed early warning index, construct a seismic acoustic electric comprehensive early warning data set, and divide it into training set, validation set and test set; the step S40 specifically comprises the following steps: S41: Normalize the collected microseismic data, acoustic emission data and electromagnetic radiation data, extract the key feature parameters of electromagnetic radiation, acoustic emission and microseismic signal according to "day" as time unit, including electromagnetic radiation intensity maximum average, electromagnetic radiation intensity average, electromagnetic radiation event number, acoustic emission intensity maximum average, acoustic emission intensity average, acoustic emission event number, microseismic maximum energy value and microseismic event average energy value; combine the field monitoring records and expert experience score to calculate the unified early warning index; S42: Build a single time series sample with daily seismic acoustic electric monitoring features and corresponding early warning index to form a seismic acoustic electric comprehensive early warning data set, of which 80% samples are divided into training set, and verified by K-fold cross-validation, and the remaining 20% samples without early warning index are used as test set; 2. The automated machine learning based seismic acoustic electro-fusion rock burst early warning method according to claim 1, characterized in that , S50: Construct a seismic acoustic electric multi-parameter comprehensive early warning module based on AutoGluon, input the key features of multi-source monitoring data, integrate regression algorithm, automatically complete feature processing, model selection and hyperparameter optimization, build a high-precision regression prediction model for impact ground pressure early warning index, use the training set to learn and fit various models, automatically search and optimize the hyperparameter combination during training, and perform iterative testing and performance evaluation on the validation set to ensure that the optimal model structure and parameter configuration are obtained, input the test set into the regression prediction model, and evaluate the model performance by RMSE index to realize automatic selection and accurate prediction of the optimal model. In the step S20, the seismic acoustic electric data collected on site is preprocessed, a seismic acoustic electric fusion prediction data set is constructed and divided, comprising the following steps: S21: The missing values and outliers of acoustic emission data and electromagnetic radiation data are filled by sliding mean value, the time steps are unified and standardized; the microseismic data with energy greater than 1.0×10 3 J events, the missing values are filled by the mean value of the day and the steps are unified; finally, the three types of data are fused into a unified structure of multi-source input data set; S22: 88% of the multi-source input data set is divided into a training set and 12% into a test set. The automatic prediction technology is used in the training stage. Based on K-fold cross-validation, the training and validation subsets are divided for model training, evaluation and hyperparameter optimization to improve the generalization ability.

3. The automated machine learning based seismic acoustic electro-fusion rock burst early warning method according to claim 1, characterized in that , In the step S30, an automatic machine learning AutoGluon-based seismic acoustic electric automatic modeling and prediction module is constructed, including the following steps: S31: Data import: The automatic machine learning AutoGluon-based seismic acoustic electric automatic modeling and prediction module is constructed, each time series of the preprocessed multi-source signal is input into the system as an independent modeling item, and the prediction range is flexibly set according to the specific task requirements to ensure that the model input meets the prediction requirements of different levels; S32: Model training: The MAEP module uses the automatic machine learning framework of AutoGluon to model using an automatic model management mechanism. The system explores a variety of model structures and hyperparameter configurations within the set time budget. For each input time series, the training set is used to fit and learn multiple prediction models, and the MASE error index on the validation set is used to sort the model performance to automatically select the best model for prediction. After the model training is completed, all models and their results are saved in a designated folder; S33: Model analysis and evaluation: The MAEP module visually compares the true value and the predicted value by image method to intuitively and qualitatively evaluate the prediction trend and correlation. At the same time, the mean absolute scaling error is used for quantitative evaluation. The smaller the MASE value, the better the prediction performance, which is used to compare the advantages and disadvantages of different models and select the best model. The test set is input into the optimal model to realize accurate prediction and early warning of future seismic acoustic electric signals.

4. The automatic machine learning based seismic acoustic emission and electrofusion rock burst early warning method according to claim 3, characterized in that , In the step S33, it is assumed that y denotes the true value, denotes the predicted value, n is the sample number, and the calculation formula of MASE is: MASE = .

5. The automated machine learning based seismic acoustic electro-fusion bumping warning method of claim 1, wherein , In the step S50, an automatic machine learning AutoGluon-based seismic acoustic electric multi-parameter comprehensive early warning module is constructed, including the following steps: S51: Data import: The automatic machine learning AutoGluon-based seismic acoustic electric multi-parameter comprehensive early warning module is constructed, and the preprocessed multi-source monitoring data is input. This data covers key feature parameters of electromagnetic radiation, acoustic emission and microseismic signals, and uses early warning index as the prediction target to unify the data format into a structured regression task and input into the system for subsequent modeling analysis; S52: Model training: The MAEW module is modeled based on the AutoGluon automatic machine learning framework. The module intelligently generates and trains multiple prediction models based on the values of multiple feature columns in the data table to predict the target column. MAEW fits and learns multiple types of algorithms through the training set to comprehensively mine the nonlinear and high-dimensional feature correlations in the data. Then, the root mean square error index on the validation set is used to evaluate and sort the performance of each model, and multiple excellent sub-models are integrated through a multi-model integration strategy to finally automatically construct a prediction model with optimal generalization ability. S53: Model analysis and evaluation: In the prediction stage, the MAEW module automatically applies the best model combination to the test set for regression prediction; in the evaluation process, the root mean square error (RMSE) is selected as the core performance indicator to reflect the accuracy of the model by measuring the overall deviation between the predicted value and the true value; the smaller the RMSE, the better the prediction effect.

6. The automated machine learning based seismic acoustic electro-fusion bumping warning method of claim 5, wherein , In the step S53, in the model analysis and evaluation process, the negative value of the RMSE is used as the score basis to sort the models, and the final output model has the optimal performance. Assuming that y represents the true value, represents the predicted value, n is the sample number, and the calculation formula of the RMSE is: 。 7. An automatic machine learning based seismic acoustic electrofusion rock burst early warning system, characterized in that The method comprises the following steps: The seismic, acoustic and electric data acquisition module is configured to construct a multi-source monitoring network covering the working face based on the field layout of the microseismic system and the acoustic and electric system, and to screen and obtain high-quality microseismic data, acoustic emission data and electromagnetic radiation data; The seismic, acoustic and electric prediction data set construction module is configured to preprocess the microseismic data, acoustic emission data and electromagnetic radiation data collected on site to form a seismic, acoustic and electric fusion prediction data set with unified time stamp and input structure, and to divide the prediction data set into training set, validation set and test set; The seismic, acoustic and electric automatic modeling and prediction module is configured to construct an automatic modeling and prediction module based on AutoGluon, take the preprocessed multi-source signals as input, realize the automatic search and optimization of various prediction model structures and hyperparameter configurations relying on the AutoGluon framework, construct a prediction system sensitive to nonlinear and multi-scale dynamic changes, use the training set to fit and learn the model, then evaluate and optimize the hyperparameters of the model through the validation set, select the model with the optimal MASE index on the validation set within a specified time, and apply it to the test set to realize accurate prediction of the seismic, acoustic and electric signals; The seismic, acoustic and electric prediction data set construction module is configured to normalize the microseismic data, acoustic emission data and electromagnetic radiation data collected on site, extract the key feature parameters of electromagnetic radiation, acoustic emission and microseismic signals according to time units, combine the artificially constructed early warning index, construct a seismic, acoustic and electric comprehensive early warning data set, and divide it into training set, validation set and test set; The seismic, acoustic and electric multi-parameter comprehensive early warning module is configured to construct a seismic, acoustic and electric multi-parameter comprehensive early warning module based on AutoGluon, take the key features of multi-source monitoring data as input, integrate regression algorithms, automatically complete feature processing, model selection and hyperparameter optimization, construct a high-precision regression prediction model for impact ground pressure early warning index, use the training set to learn and fit various models, automatically search and optimize the hyperparameter combination during the training process, and perform iterative testing and performance evaluation on the validation set to ensure that the optimal model structure and parameter configuration are obtained, input the test set into the regression prediction model, evaluate the model performance with the RMSE index, and realize automatic selection and accurate prediction of the optimal model.

8. The automated machine learning based seismic acoustic electro-fusion bumping warning system of claim 7, wherein , The seismic, acoustic and electric prediction data set construction module comprises: The data preprocessing unit is configured to perform sliding mean filling for missing values and abnormal values of acoustic emission data and electromagnetic radiation data, unify time steps, and standardize. 3 J events, missing values are filled with the mean value of the day and the step is unified; finally, the three types of data are fused into a unified structure of multi-source input data set; The data set division unit is configured to divide 88% of the multi-source input data set as a training set and 12% as a test set, and an automatic prediction technology is used in the training stage, and a K-fold cross-validation is used to divide the training and verification subsets for model training, evaluation and hyperparameter optimization to improve the generalization ability.

9. The automated machine learning based seismic acoustic electro-fusion bumping warning system of claim 7, wherein , The seismic, acoustic and electric automatic modeling and prediction module includes: The data import unit is configured to build a seismic, acoustic and electric automatic modeling and prediction module based on AutoGluon automatic machine learning, to input each time series of the preprocessed multi-source signal into the system as an independent modeling item, and to flexibly set the prediction range according to the specific task requirements to ensure that the model input meets the prediction requirements of different levels. The model training unit is configured to use the automatic machine learning framework of AutoGluon in the MAEP module to use an automatic model management mechanism for modeling; the system maximizes the exploration of various model structures and hyperparameter configurations within the set time budget; for each input time series, the system trains multiple models and sorts the model performance based on the MASE error index on the validation set to automatically select the best model for prediction; after the model training is completed, all models and their results are saved in a designated folder. The model analysis and evaluation unit is configured to visually compare the true value and the predicted value by using the image method to intuitively and qualitatively evaluate the prediction trend and correlation; at the same time, the mean absolute scaled error is used for quantitative evaluation, and the smaller the MASE value, the better the prediction performance, which is used to compare the advantages and disadvantages of different models and select the best model.

Images