Test-theory-data hybrid-driven intelligent early warning method for coal and gas outburst

Abstract

The invention discloses a test-theory-data hybrid-driven intelligent early warning method for coal and gas outburst. The method comprises the following steps: constructing a test-theory-data hybrid-driven model; based on a test-theory-data hybrid drive model, constructing a digital twin virtual model; the method comprises the following steps: deducing and feeding back underground scenes and data in real time by using a digital twin virtual model, preprocessing the data, constructing a deep learning combination model, inputting the preprocessed data into the deep learning combination model, extracting and weighting twin data features, and outputting prediction data; a multi-parameter comprehensive judgment criterion combining a threshold value method, a dynamic trend method and a comprehensive judgment method is constructed, prediction data and field factors are distinguished through the judgment criterion, the outburst risk early warning level is determined, and intelligent early warning of coal and gas outburst is completed. According to the method, existing experimental research is fully utilized, theoretical characteristics and data association of coal and gas outburst are fully mined, and the accuracy and applicability of early warning are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of coal-rock dynamic disaster monitoring and early warning, and in particular relates to an intelligent early warning method for coal and gas outburst driven by a hybrid of experiment, theory and data. Background Art

[0002] Coal and gas outburst is a special phenomenon of gas outburst. With the increase of coal mining intensity and depth, ground stress, gas pressure and gas content also increase, which makes the outburst more harmful and the early warning more difficult. The prevention and control of coal and gas outburst faces more challenges. Advance warning is the prerequisite for effective prevention of outburst, and intelligent warning is the development direction of outburst warning technology. The implementation process of existing early warning technology is to analyze the characteristics of real-time monitoring data and compare them with the characteristics of outburst precursors to determine whether there is an outburst risk. The early warning indicators and rules used are usually determined based on manual experience or on-site inspections. There are problems such as large influence of human factors, insufficient mining and utilization of historical data and precursor information, low fusion of multi-source information, relatively solidified indicators and models, insufficient self-learning and self-optimization capabilities, etc. The application effect has not fully met the ideal requirements. It is necessary to further combine digital twins, deep learning and other technologies to fully explore the theoretical characteristics and data associations of coal and gas outbursts, and improve the universality and prediction accuracy of early warning indicators and methods. The deep learning combination model can give full play to the learning advantages, extract the time series characteristics of different types of coal mine monitoring data and output them, and realize the advanced intelligent warning of gas outburst. However, there are problems such as the difficulty in obtaining field data and the lack of internal mechanism to explain the decision-making process of the model. As an emerging virtual simulation technology, digital twin technology can use the existing theoretical model and data foundation to complete the transformation of underground physical entities into virtual models, achieve the effect of virtual-real combination and synchronous simulation, realize advanced simulation and timely feedback of field conditions, integrate theoretical models and expand data samples, and provide support for further combining deep learning technology to realize intelligent early warning of outburst. However, the existing digital twin technology mostly uses theoretical modeling or data modeling methods to construct twin models, which has problems such as long development cycle and difficulty in solving. It is necessary to combine theory and data to realize the joint method of dual drive of the two. In addition, physical simulation test is a common method in the study of coal and gas outburst, which can approximate the occurrence and development process of outburst to the greatest extent under the premise of ensuring safety, provide a way to analyze the precursor information and disaster-causing effects of outburst, further derive and verify the theoretical model of outburst, and provide experimental monitoring data of outburst process. Combining physical simulation experiments to further integrate theory and data and jointly drive the construction of digital twin models can further improve the accuracy and applicability of the model and provide more effective input data for the deep learning model. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes an intelligent early warning method for coal and gas outburst driven by a hybrid experiment-theory-data. By establishing an experiment-theory-data hybrid driven model, a digital twin virtual simulation model of a coal mine underground is constructed, real-time simulation and feedback of the on-site situation are performed, and feedback data is provided to input the deep learning combination model to realize data feature extraction and prediction output. Combined with multi-parameter comprehensive judgment criteria, advanced prediction and intelligent early warning of coal and gas outbursts are realized.

[0004] To achieve the above objectives, the present invention provides an intelligent early warning method for coal and gas outburst driven by experiment, theory and data, comprising:

[0005] Constructing a hybrid experimental-theoretical-data driven model;

[0006] Based on the experimental-theoretical-data hybrid driving model, a digital twin virtual model is constructed;

[0007] Using the digital twin virtual model to deduce and feedback downhole scenes and data in real time, preprocessing the data, building a deep learning combination model, inputting the preprocessed data into the deep learning combination model, extracting and weighting twin data features, and outputting predicted data;

[0008] A multi-parameter comprehensive judgment criterion combining the threshold method, dynamic trend method and comprehensive evaluation method is constructed, and the prediction data and on-site factors are distinguished by the judgment criterion, the warning level of the outburst risk is determined, and the intelligent warning of coal and gas outburst is completed.

[0009] Optionally, constructing a hybrid experiment-theory-data driven model includes:

[0010] According to the actual mining conditions in the underground mine, the test similarity ratio is determined and physical simulation tests of coal and gas outburst are carried out;

[0011] Combined with the existing outburst-related theories and experimental results, the multi-parameter correlation law in the whole outburst process is analyzed and a comprehensive outburst theoretical model is constructed;

[0012] Collect multi-source heterogeneous data from underground sites to form real-time spatiotemporal perception information underground;

[0013] Combining physical simulation experiments on coal and gas outbursts, comprehensive theoretical models of outbursts and real-time spatiotemporal perception information underground, an experiment-theory-data hybrid driven model is constructed.

[0014] Optionally, based on the test-theory-data hybrid driven model, constructing the digital twin virtual model includes: using field and test parameters to set parameters of the digital twin model;

[0015] Use multi-source heterogeneous data from the field to perform numerical simulation analysis of the model, generate virtual representations of various field parameters, build a digital virtual model, and update it in real time based on changes in field monitoring data;

[0016] Combined with the highlighted comprehensive theoretical model content, including the changes in the underground production process and the interaction principles between different factors, a virtual-reality interaction mechanism between the digital virtual model and real-time perception information is established to complete the construction of the digital twin virtual model.

[0017] Optionally, the method of preprocessing the data includes: data supplementation, data cleaning, and data normalization;

[0018] Among them, data supplementation is used to fill in missing data forward;

[0019] The data cleaning calculation is as follows,

[0020]

[0021] Among them, x t is the data after the abnormal data value is corrected; t is the time when the data is abnormal; k is the amount of data selected before and after time t;

[0022] The data normalization calculation is as follows,

[0023]

[0024] Among them, X is the data sample, X min is the minimum value of the data, X max is the maximum value of the data, X n is the normalized target matrix.

[0025] Optionally, the preprocessed data is input into the deep learning combined model to extract and weight the twin data features, and various prediction data are output including:

[0026] The preprocessed data is input into the deep learning combined model, the convolutional neural network is used to reduce the dimension of the data and extract the local feature parameter values, the bidirectional long short-term memory neural network is used to extract the time series feature parameter values ​​of the data, the self-attention mechanism is used to analyze the feature parameter values ​​of the data and divide the weight ratio, and various types of prediction data are output through the deep learning combined model.

[0027] Optionally, the method for extracting features from the pre-processed input data in the process of reducing the dimension of the data and extracting the local feature parameter values ​​using a convolutional neural network is:

[0028]

[0029] Among them, Y (l)is the output feature, l represents the lth convolutional layer, f(·) is the activation function, n is the number of filters, and W (l),n is the filter weight, Y (l-1) is the output feature of the previous layer, b (l) For bias.

[0030] Optionally, the method for selecting the extracted features in the process of reducing the dimension of data and extracting local feature parameter values ​​using a convolutional neural network is:

[0031] X=f[αS(x)+b]

[0032] Where X is the output; f(·) is the activation function; α is the multiplicative bias; S(x) is the downsampling function; and b is the bias.

[0033] Optionally, a method for extracting the time series characteristic parameter value of data using a bidirectional long short-term memory neural network is:

[0034] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0035] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0036]

[0037] o t =σ(W ° ·[h t-1 ,x t ]+b o )

[0038] h t =o t tanh(C t )

[0039] Among them, f t ,i t , o t They are forget gate, input gate, and output gate respectively; is the new data information at time t; C t is the updated neural unit information state; x i ,h t-1 are the input and output of the neural unit at time t-1 respectively; W f , W i , W o , WC and b f , b i , b o , b C is the weight matrix and bias of the forget gate, input gate, output gate, unit information update of the neural unit; σ is the Sigmoid activation function of the neural unit; tanh is the activation function.

[0040] Optionally, the method of using the self-attention mechanism to analyze the characteristic parameter values ​​of the data and divide the weight ratio is:

[0041]

[0042] Among them, Y is the feature matrix output by the LSTM layer; M is the result of the feature matrix Y after the tanh function operation; w α is the weight matrix; α is the normalized weight matrix; A is the output vector after the weights are assigned by the attention mechanism layer; T represents the transpose of the matrix.

[0043] Optionally, a method of constructing a multi-parameter comprehensive judgment criterion combining a threshold method, a dynamic trend method, and a comprehensive evaluation method, and using the judgment criterion to distinguish the predicted data and on-site factors to determine the outstanding risk warning level includes:

[0044] The threshold method is used to calculate the over-limit warning value as follows:

[0045]

[0046] Among them, S is the over-limit warning value, n is the number of predicted data values ​​within the over-limit warning time, V n It is the predicted data value within the over-limit warning time;

[0047] The dynamic trend method is used to calculate the dynamic trend change rate as follows:

[0048]

[0049] Among them, A is the dynamic trend change rate, v 1 is the predicted data value at the previous time point, v 2 is the predicted data value at the next time point, and Δt is the time difference between the two time points;

[0050] By adopting a comprehensive evaluation method, the predicted data and on-site factors are converted into similarity coefficients and correlation degrees by mathematical methods, and then quantitatively assigned to make judgments, make quantitative evaluations and risk warnings that are in line with reality, and determine the level of prominent risk warnings.

[0051] Technical effect of the present invention: The present invention discloses an intelligent early warning method for coal and gas outburst driven by a hybrid of experiment, theory and data, and combines the existing theoretical model and data foundation in combination with the physical simulation experiment to jointly drive the construction of the digital twin model, thereby solving the problems of long development cycle and greater difficulty in solving the existing digital twin technology that uses theoretical modeling or data modeling methods to construct twin models, and at the same time improves the accuracy of the digital twin model. The present invention uses digital twin technology to realize the transformation of underground field physical entities into virtual models, achieves the effect of virtual-real combination and synchronous simulation, realizes advanced simulation and timely feedback of field conditions, integrates theoretical models and expands data samples, and provides more and more effective input data for deep learning models. The present invention uses an improved deep learning combination model to realize the extraction and prediction output of data time series features, and can distinguish different types of data types in combination with multi-parameter comprehensive judgment criteria, and further realizes the intelligent prediction and early warning level determination of coal and gas outbursts. The present invention can make full use of existing experimental research, fully explore the theoretical characteristics and data associations of coal and gas outbursts, and further improve the accuracy and applicability of early warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0053] Figure 1 It is a structural schematic diagram of the intelligent early warning method for coal and gas outburst driven by experiment-theory-data hybrid in an embodiment of the present invention;

[0054] Figure 2 The present invention provides a flowchart for implementing the intelligent early warning method for coal and gas outburst driven by a hybrid experiment-theory-data method. DETAILED DESCRIPTION

[0055] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0056] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0057] like Figure 1-2 As shown, this embodiment provides an intelligent early warning method for coal and gas outburst driven by a hybrid of experiment, theory and data, including:

[0058] Constructing a hybrid experimental-theoretical-data driven model;

[0059] Build a digital twin virtual model based on the experimental-theoretical-data hybrid driven model;

[0060] Use the digital twin virtual model to deduce and feedback underground scenes and data in real time, preprocess the data, build a deep learning combination model, input the preprocessed data into the deep learning combination model, extract and empower the twin data features, and output the predicted data;

[0061] A multi-parameter comprehensive judgment criterion combining the threshold method, dynamic trend method and comprehensive evaluation method is constructed. The judgment criteria are used to distinguish between predicted data and on-site factors, determine the risk warning level of coal and gas outbursts, and complete the intelligent warning of coal and gas outbursts.

[0062] Specifically, by establishing an experimental-theory-data hybrid driven model, a digital twin virtual simulation model of underground coal mines is constructed, real-time simulation and feedback of on-site conditions are carried out, feedback data is provided as input to the deep learning combination model, data feature extraction and prediction output are realized, and multi-parameter comprehensive judgment criteria are combined to achieve advanced prediction and intelligent early warning of coal and gas outbursts.

[0063] Furthermore, the construction of the experiment-theory-data hybrid driven model includes:

[0064] According to the actual mining conditions in the underground mine, the test similarity ratio is determined and physical simulation tests of coal and gas outburst are carried out;

[0065] Combined with the existing outburst-related theories and experimental results, the multi-parameter correlation law in the whole outburst process is analyzed and a comprehensive outburst theoretical model is constructed;

[0066] Collect multi-source heterogeneous data from underground sites to form real-time spatiotemporal perception information underground;

[0067] Combining physical simulation experiments on coal and gas outbursts, comprehensive theoretical models of outbursts and real-time spatiotemporal perception information underground, an experiment-theory-data hybrid driven model is constructed.

[0068] Specifically, the actual mining conditions at the underground site are investigated, including the on-site geological conditions, stress state, and tunnel settings. An appropriate similarity ratio is set, and the existing test system is used to carry out physical simulation tests of coal and gas outbursts under dynamic load disturbance conditions that restore the on-site conditions. The test phenomena and data results are collected. The test process includes device installation, outburst incubation and induction, data collection and processing. Combined with the existing outburst induction, development, and disaster-causing theories and physical simulation test results, the evolution laws of the multi-physical fields of the coal body during the whole process of the outburst are analyzed, and the multi-parameter precursor response characteristics and their relationship with the outburst hazard during the outburst incubation process are studied. Based on the dynamic instability conditions during the outburst process, a dynamic model of outburst instability is constructed. Remote sensing detection technology, downhole sensors, and parameter inversion methods are used to collect multi-source heterogeneous data at the underground site, including geological measurement data, real-time monitoring data, and parameter inversion results. Combining the results of physical simulation tests, theoretical models, and field multi-source heterogeneous data, a test-theory-data hybrid driven model is constructed, in which the simulation test can verify the accuracy of the field multi-source heterogeneous data, help build the theoretical model framework and verify its applicability. The theoretical model can guide the design of simulation tests and the collection and optimization of field multi-source heterogeneous data. The field multi-source heterogeneous data can serve as a reference to support the implementation of simulation tests and help the theoretical model to adjust parameters and correct the model. The three are integrated to improve each other's reliability.

[0069] Further, based on the test-theory-data hybrid driven model, constructing a digital twin virtual model includes: using field and test parameters to set parameters of the digital twin model;

[0070] Use multi-source heterogeneous data from the field to perform numerical simulation analysis of the model, generate virtual representations of various field parameters, build a digital virtual model, and update it in real time based on changes in field monitoring data;

[0071] Combined with the highlighted comprehensive theoretical model content, including the changes in the underground production process and the interaction principles between different factors, a virtual-reality interaction mechanism between the digital virtual model and real-time perception information is established to complete the construction of the digital twin virtual model.

[0072] Specifically, the physical entities of the digital twin model include underground production personnel, tunneling equipment, production environment, and production process; the virtual model is the basis of mapping, which digitally describes the underground scene and collects historical data, real-time data, and operating status; twin data is the driving chain of the model, including physical body data, virtual body data, service data, knowledge data, and fused derivative data; the connection channel provides a way for each part to collaborate and interact, maintains the consistency of the physical entity and virtual model of the digital twin system, and enables each part to continuously interact and update; the service core can realize the ultimate function of the model, including online perception and status analysis of underground production scenes, and feedback to the twin database for monitoring and early warning.

[0073] Furthermore, the underground scenes collected for real-time deduction and feedback of the digital twin model include on-site environment, equipment layout, personnel arrangement, and operation process, and the real-time data include environmental parameters, coal mining status parameters, and operation data; the methods for preprocessing the data include: data supplementation, data cleaning, and data normalization; among them, the role of data supplementation is to avoid the impact of missing samples on the sequence integrity of the data and improve the prediction effect. The method of forward filling of missing data can be used, that is, the missing value is filled with the previous non-missing value.

[0074] The purpose of data cleaning is to remove outliers and then use data supplementation methods to fill them or directly correct them with reasonable values. This process uses the nearest mean method. The data cleaning calculation is as follows:

[0075]

[0076] Among them, x t is the data after the abnormal data value is corrected; t is the time when the data is abnormal; k is the amount of data selected before and after time t;

[0077] The role of data normalization is to eliminate the influence of different dimensions and orders of magnitude between parameters and improve the convergence speed of the model. The normalization calculation is as follows:

[0078]

[0079] Among them, X is the data sample, X min is the minimum value of the data, X max is the maximum value of the data, X n is the normalized target matrix.

[0080] Furthermore, the preprocessed data is input into the deep learning combined model, the twin data features are extracted and weighted, and various types of prediction data are output, including: the preprocessed data is input into the deep learning combined model, a convolutional neural network is used to reduce the dimension of the data and extract local feature parameter values, a bidirectional long short-term memory neural network is used to extract the time series feature parameter values ​​of the data, a self-attention mechanism is used to analyze the feature parameter values ​​of the data and divide the weight ratio, and various types of prediction data are output through the deep learning combined model.

[0081] Specifically, the deep learning combination model is a deep learning model composed of a convolutional neural network (CNN), a bidirectional long short-term memory neural network (BiLSTM), and a self-attention mechanism (SA). The three structures are used to extract local features of data, extract temporal features of data bidirectionally, and give greater weights to features that have a greater impact on the prediction results to improve the prediction accuracy.

[0082] After the preprocessed data is input into the combined model, the convolutional neural network (CNN) is used to reduce the dimension of the data and extract the local feature parameter values. The CNN structure includes input layer, convolution layer, pooling layer, fully connected layer and output layer. The function of the convolution layer is to extract the features of the input layer data. Assume that there are n convolution layers in the lth layer. (l) filters with weights W (l) ∈R 3×3×n(l－1) , the size is 3×3, where n (l－1) is the number of filters in the previous layer. The filter traverses the entire input feature map with a step of [1,1], and the output feature is:

[0083]

[0084] Among them, Y (l) is the output feature, l represents the lth convolutional layer, f(·) is the activation function, n is the number of filters, and W (l),n is the filter weight, Y (l-1) is the output feature of the previous layer, b (l) For bias.

[0085] The role of the pooling layer is to select the features extracted by the convolutional layer, reduce the overfitting of the model and improve its generalization ability. The expression is:

[0086] X=f[αS(x)+b]

[0087] Where X is the output; f(·) is the activation function; α is the multiplicative bias; S(x) is the downsampling function; and b is the bias.

[0088] The fully connected layer then performs a nonlinear combination of the features filtered by the pooling layer;

[0089] The bidirectional long short-term memory neural network (BiLSTM) is used to extract the time series feature parameter values ​​of the data. The time series features include proximity, periodicity and trend. The key to BiLSTM is the information state of the neural unit. The information flow of the neural unit is controlled by three logic gates (forget gate, input gate, output gate). The processing process is shown as follows:

[0090] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0091] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0092]

[0093] o t =σ(W ° ·[h t-1 ,x t ]+b o )

[0094] h t =o t tanh(C t )

[0095] where f t ,i t , o t They are forget gate, input gate, and output gate respectively; is the new data information at time t; C t is the updated neural unit information state; x i ,h t-1 are the input and output of the neural unit at time t-1 respectively; W f , W i , W o , W C and b f , b i , b o , b C is the weight matrix and bias of the forget gate, input gate, output gate, and unit information update of the neural unit; σ is the Sigmoid activation function of the neural unit; tanh is the activation function;

[0096] Combined with the content of the theoretical model, according to the impact of different types and time data on the prediction results, the self-attention mechanism is used to analyze the characteristic parameter values ​​of the data and divide the weight ratio. Finally, various types of prediction data are output through the deep learning combination model. The method of using the self-attention mechanism to analyze the characteristic parameter values ​​of the data and divide the weight ratio is:

[0097]

[0098] Among them, Y is the feature matrix output by the LSTM layer; M is the result of the feature matrix Y after the tanh function operation; w α is the weight matrix; α is the normalized weight matrix; A is the output vector after the weights are assigned by the attention mechanism layer; T represents the transpose of the matrix.

[0099] Finally, in order to verify the prediction effect of the model, the root mean square error (RMSE), mean absolute error (MAE) and mean relative percentage error (MAPE) are selected to evaluate the quality of the model. The smaller the three indicators are, the better the prediction effect of the model is. The calculation formula is as follows:

[0100]

[0101]

[0102] Among them, y i Represents the actual value of the data; represents the predicted data; m represents the number of samples.

[0103] Furthermore, a multi-parameter comprehensive judgment criterion combining a threshold method, a dynamic trend method, and a comprehensive evaluation method is constructed, and the prediction data and on-site factors are distinguished by using the judgment criterion. The method for determining the early warning level of the outstanding risk includes:

[0104] Among them, the threshold method is suitable for prediction data with little volatility. According to the provisions of the "Coal Mine Safety Regulations" and other normative documents and the actual situation on site, the over-limit warning threshold, over-limit warning time and outburst risk warning level judgment standard of the monitoring parameters are determined. The predicted data value output by the above steps is compared with the over-limit warning threshold. When the predicted data value exceeds the over-limit warning threshold, the occurrence time is immediately recorded. When the occurrence time reaches the over-limit warning time, it is determined that there is an outburst risk. At the same time, the over-limit warning value is calculated to determine the outburst risk warning level. The calculation formula of the over-limit warning value is as follows:

[0105]

[0106] Among them, S is the over-limit warning value, n is the number of predicted data values ​​within the over-limit warning time, V n It is the predicted data value within the over-limit warning time;

[0107] The dynamic trend method is suitable for forecast data with volatility and can reflect the danger of parameter change trend over a period of time. Based on this, combined with the over-limit warning triggering conditions in the threshold method, the risk level of the outburst can be judged. The implementation process is: collect the changes in the forecast data values ​​within a certain period of time, and obtain its dynamic trend change rate, which is compared with the trend change rate of such parameters before the outburst in the existing theoretical model, to predict the probability of the outburst and determine the outburst risk warning level. The dynamic trend change rate calculation formula is:

[0108]

[0109] Among them, A is the dynamic trend change rate, v 1 is the predicted data value at the previous time point, v 2 is the predicted data value at the next time point, and Δt is the time difference between the two time points;

[0110] The comprehensive evaluation method is suitable for prediction data and field factors with unclear boundaries and difficult quantitative analysis. It can be combined with qualitative analysis to convert the prediction data and field factors into quantitative values ​​such as similarity coefficient and correlation by mathematical methods, and then make judgments to make quantitative evaluations and risk warnings that are closer to reality, and determine the level of early warning of coal and gas outburst risks. The methods that can be used in early warning of coal and gas outbursts include hierarchical analysis method and fuzzy comprehensive evaluation method.

[0111] The prediction data and on-site factors output by the digital twin model and the neural network model are analyzed using multi-parameter comprehensive judgment criteria, and the risk warning level of coal and gas outbursts is divided and output, ultimately achieving intelligent real-time warning of coal and gas outbursts.

[0112] The present invention discloses an intelligent early warning method for coal and gas outburst driven by a hybrid of experiment, theory and data. The existing theoretical model and data foundation are integrated in combination with physical simulation experiments to jointly drive the construction of a digital twin model, solving the problems of long development cycle and greater difficulty in solving existing digital twin technologies that use theoretical modeling or data modeling methods to construct twin models, and at the same time improving the accuracy of the digital twin model. The present invention uses digital twin technology to realize the transformation of underground field physical entities into virtual models, achieves the effect of virtual-real combination and synchronous simulation, realizes advanced simulation and timely feedback of field conditions, integrates theoretical models and expands data samples, and provides more and more effective input data for deep learning models. The present invention uses an improved deep learning combined model to realize the extraction and prediction output of data time series features, and can distinguish different types of data types in combination with multi-parameter comprehensive judgment criteria, and further realizes the intelligent prediction and early warning level determination of coal and gas outbursts. The present invention can make full use of existing experimental research, fully explore the theoretical characteristics and data associations of coal and gas outbursts, and further improve the accuracy and applicability of early warnings.

[0113] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. Experiment-theory-data hybrid driven coal and gas outburst intelligent early warning method, It is characterized in that include: Constructing a hybrid experimental-theoretical-data driven model; Based on the experimental-theoretical-data hybrid driving model, a digital twin virtual model is constructed; Using the digital twin virtual model to deduce and feedback downhole scenes and data in real time, preprocessing the data, building a deep learning combination model, inputting the preprocessed data into the deep learning combination model, extracting and weighting twin data features, and outputting predicted data; A multi-parameter comprehensive judgment criterion combining the threshold method, dynamic trend method and comprehensive evaluation method is constructed, and the prediction data and on-site factors are distinguished by the judgment criterion, the warning level of the outburst risk is determined, and the intelligent warning of coal and gas outburst is completed.

2. The test-theory-data hybrid driven coal and gas outburst intelligent early warning method as claimed in claim 1, It is characterized in that Building a hybrid experiment-theory-data driven model includes: According to the actual mining conditions in the underground mine, the test similarity ratio is determined and physical simulation tests of coal and gas outburst are carried out; Combined with the existing outburst-related theories and experimental results, the multi-parameter correlation law in the whole outburst process is analyzed and a comprehensive outburst theoretical model is constructed; Collect multi-source heterogeneous data from underground sites to form real-time spatiotemporal perception information underground; Combining physical simulation experiments on coal and gas outbursts, comprehensive theoretical models of outbursts and real-time spatiotemporal perception information underground, an experiment-theory-data hybrid driven model is constructed.

3. The test-theory-data hybrid driven coal and gas outburst intelligent early warning method as claimed in claim 2, It is characterized in that Based on the test-theory-data hybrid driven model, constructing a digital twin virtual model includes: using field and test parameters to set parameters of the digital twin model; Use multi-source heterogeneous data from the field to perform numerical simulation analysis of the model, generate virtual representations of various field parameters, build a digital virtual model, and update it in real time based on changes in field monitoring data; Combined with the highlighted comprehensive theoretical model content, including the changes in the underground production process and the interaction principles between different factors, a virtual-reality interaction mechanism between the digital virtual model and real-time perception information is established to complete the construction of the digital twin virtual model.

4. The test-theory-data hybrid driven coal and gas outburst intelligent early warning method as claimed in claim 1, It is characterized in that The method of preprocessing the data includes: data supplementation, data cleaning, and data normalization; Among them, data supplementation is used to fill in missing data forward; The data cleaning calculation is as follows, Among them, x t is the data after the abnormal data value is corrected; t is the time when the data is abnormal; k is the amount of data selected before and after time t; The data normalization calculation is as follows, Among them, X is the data sample, X min is the minimum value of the data, X max is the maximum value of the data, X n is the normalized target matrix.

5. The test-theory-data hybrid driven coal and gas outburst intelligent early warning method as claimed in claim 1, It is characterized in that The preprocessed data is input into the deep learning combined model to extract and weight the twin data features, and various prediction data are output including: The preprocessed data is input into the deep learning combined model, the convolutional neural network is used to reduce the dimension of the data and extract the local feature parameter values, the bidirectional long short-term memory neural network is used to extract the time series feature parameter values ​​of the data, the self-attention mechanism is used to analyze the feature parameter values ​​of the data and divide the weight ratio, and various types of prediction data are output through the deep learning combined model.

6. The test-theory-data hybrid driven coal and gas outburst intelligent early warning method as claimed in claim 5, It is characterized in that The method of extracting features from the pre-processed input data in the process of reducing the dimension of the data and extracting the local feature parameter values ​​using a convolutional neural network is as follows: Among them, Y (l) is the output feature, l represents the lth convolutional layer, f(·) is the activation function, n is the number of filters, and W (l),n is the filter weight, Y (l-1) is the output feature of the previous layer, b (l) For bias.

7. The test-theory-data hybrid driven coal and gas outburst intelligent early warning method as claimed in claim 6, It is characterized in that The method for selecting the extracted features in the process of reducing the dimension of data and extracting local feature parameter values ​​using convolutional neural networks is: X=f[αS(x)+b] Where X is the output; f(·) is the activation function; α is the multiplicative bias; S(x) is the downsampling function; and b is the bias.

8. The test-theory-data hybrid driven coal and gas outburst intelligent early warning method as claimed in claim 5, It is characterized in that The method of extracting the time series characteristic parameter value of data using a bidirectional long short-term memory neural network is: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W ° ·[h t-1 ,x t ]+b o ) h t =o t ·tanh(C t ) Among them, f t ,i t , o t They are forget gate, input gate, and output gate respectively; is the new data information at time t; C t is the updated neural unit information state; x i ,h t-1 are the input and output of the neural unit at time t-1 respectively; W f , W i , W o , W C and b f , b i , b o , b C is the weight matrix and bias of the forget gate, input gate, output gate, unit information update of the neural unit; σ is the Sigmoid activation function of the neural unit; tanh is the activation function.

9. The test-theory-data hybrid driven coal and gas outburst intelligent early warning method as claimed in claim 5, It is characterized in that The method of using the self-attention mechanism to analyze the characteristic parameter values ​​of the data and divide the weight ratio is: Among them, Y is the feature matrix output by the LSTM layer; M is the result of the feature matrix Y after the tanh function operation; w α is the weight matrix; α is the normalized weight matrix; A is the output vector after the weights are assigned by the attention mechanism layer; T represents the transpose of the matrix.

10. The test-theory-data hybrid driven coal and gas outburst intelligent early warning method according to claim 1, It is characterized in that A method for constructing a multi-parameter comprehensive judgment criterion combining a threshold method, a dynamic trend method, and a comprehensive evaluation method, and using the judgment criterion to distinguish the predicted data and on-site factors to determine the level of early warning of outstanding risks includes: The threshold method is used to calculate the over-limit warning value as follows: Among them, S is the over-limit warning value, n is the number of predicted data values ​​within the over-limit warning time, V n It is the predicted data value within the over-limit warning time; The dynamic trend method is used to calculate the dynamic trend change rate as follows: Among them, A is the dynamic trend change rate, v 1 is the predicted data value at the previous time point, v 2 is the predicted data value at the next time point, and Δt is the time difference between the two time points; By adopting a comprehensive evaluation method, the predicted data and on-site factors are converted into similarity coefficients and correlation degrees by mathematical methods, and then quantitatively assigned to make judgments, make quantitative evaluations and risk warnings that are in line with reality, and determine the level of prominent risk warnings.

Images