Intelligent fusion classification and grading advanced early warning method for major disaster of coal mine

Abstract

The invention discloses an intelligent fusion classification and grading advanced early warning method for major disasters of a coal mine. Firstly, coal mine safety monitoring data are collected to construct a data set; secondly, an Angle-Transform model and an E2GAN model are trained, abnormal values and missing values in the data are recognized and interpolated, and data preprocessing is achieved; and thirdly, a new multi-index fusion prediction model is introduced in combination with a BiLSTM and MMOE multi-task learning network, and the future trend of each index can be synchronously predicted after training optimization. And then, improving an AdaTT network to realize a multi-disaster risk fusion early warning model, and optimizing training times and an early warning threshold value by using an F1 index. And inputting real-time monitoring and fusion prediction data into the optimized early warning model to obtain the risk probability of each disaster. And finally, in combination with the optimal early warning threshold and the risk probability, providing a multi-disaster grading early warning method, and realizing grading identification of different disaster risks. According to the method, disaster risks such as gas, fire disasters, dust and roofs can be effectively warned, risk information is provided in advance, and the method is of great significance to advanced management and control of multiple disaster risks of a coal mine.

Description

Technical Field

[0001] The present invention relates to the technical field of coal mine disaster early warning, and in particular to an intelligent fusion classification and grading advance early warning method for major coal mine disasters. Background Art

[0002] As shallow coal resources are gradually depleted, the increase in mining depth has led to more complex geological conditions. The risks of accidents such as gas, fire, dust, roof, water, rock burst, coal and gas outbursts are becoming increasingly severe, causing serious property losses and casualties. Comprehensive research on the mutual influence of various indicators, the risk coupling relationship between various disasters, and the risk information of different indicators before the occurrence of various disaster accidents is an important prerequisite for the prediction and early warning of various disaster risks.

[0003] At present, the traditional early warning methods for disaster risks such as gas, coal dust, fire, roof, water, rock burst, coal and gas outburst are mostly based on the time series precursor characteristics of single factor indicators and independent judgment and warning of data range, such as trend method, critical value method or machine learning models for independent judgment of each disaster, such as CNN, LSTM, etc. These methods ignore the mutual influence and correlation between each disaster and each indicator. Due to the complex environmental conditions and coupling processes of coal mine disasters, unclear influencing factors and evolution processes, this early warning method that only considers a single indicator cannot quantify and accurately express the disaster characteristics of each disaster, ignores the mutual influence relationship between each disaster and each indicator, lacks coordinated monitoring and application of multiple disasters and multiple indicators, and cannot accurately integrate and identify the danger of each disaster risk, resulting in the inability to timely and effectively identify and control risks and hidden dangers, which is easy to cause the underreporting or false reporting of the judgment of disaster risks such as gas, coal dust, fire, roof, water, rock burst, coal and gas outburst, and reduce the accuracy and reliability of risk warning. Summary of the invention

[0004] Purpose of the invention: In view of the above problems, the purpose of the present invention is to provide an intelligent fusion classification and grading advance warning method for major coal mine disasters, which can integrate factors such as gas, coal dust, fire, roof, water, impact ground pressure, coal and gas outburst and other disaster risks for advance warning, thereby improving the accuracy and reliability of the warning.

[0005] Technical solution: A method for intelligent fusion classification and grading of coal mine major disasters, including the following steps:

[0006] Step 1: Data preprocessing;

[0007] Collect a large amount of safety monitoring time series data in coal mines, including time series data of gas concentration, carbon monoxide concentration, coal dust concentration, fully mechanized mining resistance, surrounding rock separation, anchor cable stress, temperature, wind speed, water inflow, etc. as the initial data set, and use deep learning models to identify and interpolate abnormal data in the time series data to achieve data preprocessing;

[0008] Step 2: Fusion prediction;

[0009] A multi-indicator fusion prediction model is established based on a bidirectional long short-term memory network combined with an improved MMoE algorithm. The preprocessed data is input into the multi-indicator fusion prediction model. By learning the precursor trend characteristics of each indicator data and the correlation between indicators, the future development trends of different indicators are obtained simultaneously.

[0010] Step 3: Fusion warning;

[0011] A multi-disaster risk fusion warning model based on the improved AdaTT multi-indicator fusion prediction model is established. The original data and predicted data are input into the fusion warning model. The current and future risk probabilities of each disaster are output synchronously by using the fusion analysis of the data precursor characteristics and the relationship between disasters. The optimal model is determined by the evaluation indicators, and the warning thresholds of each disaster risk are obtained by combining the accuracy distribution in the test set.

[0012] Step 4: Classification and identification;

[0013] The current and future risk probabilities of various disasters obtained by the fusion warning model are combined with the risk warning thresholds of various disasters to establish a disaster risk classification identification and warning method that integrates current and future risk conditions. This method is used to carry out comprehensive risk identification and warning of disasters such as gas, fire, dust, roof, water, rock burst, coal and gas outbursts.

[0014] Furthermore, in step 1, the collected abnormal data is divided into training set, validation set and test set, and Anomaly-Transformer and E 2 GAN establishes a data pre-training model and performs training and testing. 2 GAN identifies and interpolates outliers and missing values ​​in the time series data of gas concentration, carbon monoxide concentration, coal dust concentration, comprehensive mining resistance, surrounding rock delamination, anchor cable stress, temperature, wind speed, water inflow, etc. collected on site.

[0015] Furthermore, in step 2, the method for establishing a multi-index fusion prediction model includes:

[0016] Assume that all the input index data is x, and the index prediction task is K = {k1, k2, ...k n}, the number of expert networks is N, and the long-term dependency features of each indicator are extracted through a bidirectional long short-term memory network. The calculation method is as follows:

[0017]

[0018] in, is the candidate information memory cell of index k at time t, is the information memory cell of index k at time t, is the information output of the k index at time t; then the long-term dependency features of different indicators obtained by bidirectional long short-term memory network learning are spliced ​​to obtain the splicing result X c :

[0019] X c =[H1,H2,...,H k ];

[0020] Among them, H k It is the result output by the bidirectional long short-term memory network of the kth index;

[0021] Finally, the spliced ​​matrix is ​​input into the multi-task learning model MMoE for fusion calculation, and the prediction results of each indicator are output synchronously after the task tower; the method is as follows:

[0022]

[0023] Among them, G k (x) is the gate control unit of task k, W gk To learn the parameter weights, f k (x) is the output of task k after learning through the parameter-sharing expert network, m k For the mission tower network, y k It is the prediction result of k tasks obtained by fusion analysis of all indicator data.

[0024] Furthermore, in step three, the method for determining each disaster warning threshold includes:

[0025] Assume that the input data of each disaster risk monitoring indicator is x, and its warning task is T = {t1, t2, ..., t n}, the number of fusion unit stacking layers is L = {l1, l2, ..., l n}, the t-th task fusion module in the l-th layer is represented as Then the fusion output of the t-th warning task at the l-th layer is for:

[0026]

[0027] The calculation process of each fusion network block FU is to first integrate the outputs of multiple expert networks of key indicators through a linear model. The formula is as follows:

[0028]

[0029] in, After superimposing the matrix for each expert, is the parameter vector, NELF is the output of multiple experts of key indicators after linear fusion; then, the gating network is used to learn the correlation between other indicators and key indicators, and all the expert network outputs are fused to obtain the influence weight of each factor on the evolution of disaster risk. The formula is as follows:

[0030]

[0031] Finally, the gating network results are added to the results of the key indicator expert network fusion to obtain a data fusion block for the disaster risk warning task. The calculation formula is as follows:

[0032]

[0033] Optimally, in step three, the optimal model of the warning model is determined by the optimal F1 value method. The model determination method based on the threshold of F1 value maximization is used to obtain the number of model iterations when F1 is optimal in the iterative training process and determine its corresponding threshold in combination with the accuracy distribution of the optimal model in the test set. F1 is the harmonic mean of precision and recall. The values ​​of precision and recall are further modified by adjusting the threshold so that the F1 value reaches the maximum, thereby determining the warning thresholds for each disaster risk.

[0034] Due to the change of threshold in the iterative process, the model obtained in one iteration can get multiple F1 values. Here, based on the interval of 0.01, the model with the largest F1 is selected as the model of this iteration. Compared with other iterations, the corresponding threshold is equivalent to the warning threshold of the iterative model, that is, each iteration selects the threshold with the best F1 under different thresholds to calculate, which is used as the optimal F1 value of this iteration, and then selects the optimal model from multiple iterations, and chooses the best from the best.

[0035] The best way to determine the F1 value is:

[0036] Suppose the precision is P. The precision P is the ratio of the number of results that are actually marked as warnings among the results that are output as warnings by the model to the number of results that are output as warnings by the model. The calculation is as follows:

[0037]

[0038] Among them, TP represents the number of samples that are actually warning and predicted as warning, and FP represents the number of samples that are actually not warning but predicted as warning;

[0039] Suppose the recall rate is R, which refers to the proportion of samples whose predicted value is warning and whose true value is also warning to all samples whose true value is warning:

[0040]

[0041] Among them, FN represents the number of samples that are actually warning but predicted as no warning, and the calculation method of F1 value is as follows:

[0042]

[0043] Furthermore, in step four, the fusion warning model is used to calculate the risk probabilities of disasters such as gas, fire, coal dust, roof, water, rock burst, coal and gas outbursts at the current time point and in the future time period. If the disaster risk probability does not exceed the threshold in both time periods, no warning will be issued. If the disaster risk probability at the current time point exceeds the threshold and the disaster risk warning probability in the future time period also exceeds the threshold, a first-level warning will be issued for the corresponding disaster, and a second-level warning will be issued for other situations.

[0044] The threshold is comprehensively considered from the distribution of the fusion warning F1 value and the accuracy of different threshold selections in the test set.

[0045] Beneficial effects: Compared with the prior art, the advantages of the present invention are: the present invention combines deep learning to integrate early warning of disaster risks such as gas, fire, dust, roof, water, rock burst, coal and gas outburst, and the fusion analysis and prediction of various monitoring index data, comprehensively considering the mutual influence rules between various indicators, the risk coupling relationship between various disasters, and the precursor characteristic information of different indicators before the occurrence of various disaster accidents to carry out advance warning of disaster risks such as gas, coal dust, fire, roof, water, rock burst, coal and gas outburst. The present invention first preprocesses the original data sequence, and the subsequent prediction and warning provides complete and high-quality sequence data, and then integrates all indicators for prediction. The prediction model shares data information between multiple tasks by sharing parameters, effectively learns the precursor trend characteristics of different indicator data and the correlation with each prediction task, has high data prediction accuracy, and provides important future data support for subsequent multi-disaster risk integration warning. Finally, the original data and future data are simultaneously input into the fusion warning model, which comprehensively learns the relationship between the precursor characteristics of historical data of different indicators, between different indicators and various disaster risk warning tasks, and between different warning tasks, which can improve the fusion analysis ability of the warning task, so that the identification of various disaster risks has a higher accuracy rate. In addition, the present invention uses the risk probabilities of current and future disaster risks such as gas, coal dust, fire, roof, water, rock burst, coal and gas outburst obtained by the fusion warning model to carry out graded warnings for different disaster risks, so as to better ensure the accurate identification of various disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the present invention;

[0047] Figure 2 Flowchart for data preprocessing training and application;

[0048] Figure 3 This is the architecture diagram of the improved MMoE multi-indicator fusion prediction model;

[0049] Figure 4 This is the architecture diagram of the improved AdaTT coal mine multi-hazard fusion early warning model. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0051] A method for intelligent fusion classification and grading of coal mine major disasters in advance warning, such as Figure 1 As shown, the following steps are included:

[0052] Step 1: Data preprocessing;

[0053] The present invention uses Anomaly-Transformer and E 2 GAN identifies and interpolates outliers and missing values ​​in the time series data of gas concentration, carbon monoxide concentration, coal dust concentration, fully mechanized mining resistance, surrounding rock separation, anchor cable stress, temperature, wind speed, water inflow, etc. collected on site. The Anomaly-Transformer model realizes unsupervised time series outlier detection through the abnormal attention mechanism and the maximum and minimum association strategy. The main idea of ​​the abnormal attention mechanism is to use the learnable Gaussian kernel and self-attention mechanism to learn the prior association of features in relative time and the sequence association characteristics of the original sequence respectively, so as to obtain more information associations when outliers appear in the sequence, so as to calculate the abnormal score of each monitoring value in the time series. The model has a good performance in outlier identification. E 2 The GAN model is an end-to-end generative adversarial network model for estimating missing values ​​in multivariate time series. Its main structure consists of a generator that reconstructs the complete sequence by mapping the low-dimensional vector after adding noise to the original sequence, and a discriminator that distinguishes the reconstructed sequence from the real sequence. Its core is to use the GURI module to ensure that the time lag phenomenon that the influence of past samples on the current moment is weakened as time goes by, and to consider the impact of irregular sample missing on time series samples, so as to reduce the dimension of incomplete time series, reconstruct, and distinguish between real samples and reconstructed samples. After multiple rounds of iterative training of the generator and the discriminator, when the discriminator cannot distinguish between the original sample and the reconstructed sample, the corresponding value of the reconstructed sample can be used to fill the missing value in the original sample. Its model architecture, training and verification process are as follows: Figure 2 shown.

[0054] Step 2: Fusion prediction and early warning;

[0055] a) Deep learning fusion prediction model based on improved MMoE;

[0056] The present invention targets disaster monitoring indicators such as gas, fire, dust, roof, water, rock burst, coal and gas outburst, and can more accurately predict the future trend of each indicator by comprehensively analyzing the coupling evolution laws between them. The multi-task prediction model can simultaneously input continuous historical data (length n) of multiple indicator data, and then fuse and analyze the correlation between the indicators in the hidden layer, comprehensively consider the mutual influence between different indicators, and finally synchronously output the prediction data of all indicators (length m). MMoE is a multi-task learning model with parameter sharing as the base. Its basic idea is to learn the correlation between input data and different tasks through multiple shared expert networks and gated units, and then output the predicted values ​​of different tasks through a task tower composed of a feedforward neural network. However, the original MMoE model does not extract the features of the internal relationship of long time series well. Adding a bidirectional long short-term memory network before sharing parameters can effectively obtain the evolution relationship between the original sequences. Assume that all the input indicator data is x, and the indicator prediction task is K={k1, k2,…k n}, the number of expert networks is N, and the long-term dependency features of each indicator are extracted through a bidirectional long short-term memory network. The calculation method is as follows:

[0057]

[0058] in, is the candidate information memory cell of index k at time t, is the information memory cell of index k at time t, is the information output of the k index at time t; then the long-term dependency features of different indicators obtained by bidirectional long short-term memory network learning are spliced ​​to obtain the splicing result X c :

[0059] X c =[H1,H2,…,H k ];

[0060] Among them, H k is the result of the kth indicator output by the bidirectional long-short time network;

[0061] Finally, the spliced ​​matrix is ​​input into the multi-task learning model MMoE for fusion calculation, and the prediction results of each indicator are output synchronously after the task tower; the fusion calculation method of MMoE is as follows:

[0062]

[0063] Among them, G k (x) is the gate control unit of task k, W gkThe parameter weights are learned for it. k (x) is the output of task k after learning through the parameter-sharing expert network, m k For the mission tower network, y k is the k-task prediction result obtained by fusion analysis of all indicator data. The improved model architecture is as follows Figure 3 shown.

[0064] b) Deep learning fusion prediction model based on improved AdaTT;

[0065] The fusion warning model is to synchronously input multiple indicator data, use the data fusion module and the multi-disaster risk warning task fusion module to learn the correlation between the precursor characteristics of each indicator historical data, between each indicator and each disaster risk warning, and between the evolution process of each disaster risk, and synchronously output the fusion warning results of different disasters. However, different indicators have different response levels to each disaster risk. Therefore, the present invention selects the gas concentration, carbon monoxide concentration, dust concentration, comprehensive mining resistance, and water inflow of the working face as the key factors for warning of disaster risks such as gas, fire, coal dust, roof, water, rock burst, coal and gas outburst, etc., and surrounding rock separation, anchor stress, temperature, wind speed, long observation hole, water level, etc. as auxiliary factors. Adatt is a multi-task adaptive deep fusion network model, which is constructed by specific warning tasks and optional shared fusion blocks at multiple levels. The main idea is to learn the regular knowledge of specific indicator data by assigning different expert networks to different warning tasks, and to fuse data to data and task to task by using residual mechanism and gating mechanism. These units adaptively learn the associated knowledge and rules between different data and specific warning tasks, thereby effectively modeling the relationship between warning tasks and realizing knowledge fusion and sharing. By superimposing multiple layers of fusion modules, the warning results of specific warning tasks under the influence of precursor characteristics of historical data of different indicators and other warning tasks can be learned from a large amount of data.

[0066] Assume that the input data of each disaster risk monitoring indicator is x, and its warning task is T = {t1, t2, ..., t n}, the number of fusion unit stacking layers is L = {l1, l2, ..., l n}, the t-th task fusion module in the l-th layer is represented as Then the fusion output of the t-th warning task at the l-th layer is for:

[0067]

[0068] The calculation process of each fusion network block FU is to first integrate the outputs of multiple expert networks of key indicators through a linear model. The formula is as follows:

[0069]

[0070] in After superimposing the matrix for each expert, is the parameter vector, and NELF is the output of multiple experts of key indicators after linear fusion. Then, the gating network is used to learn the correlation between other indicators and key indicators, and all the outputs of expert networks are fused to obtain the influence weight of each factor on the evolution of disaster risk. The formula is as follows:

[0071]

[0072] Finally, the gating network results are added to the results of the key indicator expert network fusion to obtain a data fusion block for the disaster risk warning task. The calculation formula is as follows:

[0073]

[0074] The overall architecture of the model is as follows Figure 4 shown.

[0075] In order to better select the field application model, the present invention uses a model determination method based on the threshold of F1 value maximization to obtain the number of model iterations when F1 is optimal in the iterative training process and determine the corresponding threshold value in combination with the accuracy distribution of the optimal model in the test set. F1 is the harmonic mean of precision and recall. The values ​​of precision and recall are further modified by adjusting the threshold so that the F1 value reaches the maximum, thereby determining the threshold value for each disaster risk warning. The precision rate P refers to the number of results that are actually marked as warnings in the model output as warnings, which accounts for the proportion of the number of model output as warning results, and is calculated as shown in the following formula:

[0076]

[0077] TP is the number of samples that are actually warnings and predicted as warnings, and FP is the number of samples that are actually not warnings but predicted as warnings. The recall rate R refers to the proportion of samples whose predicted values ​​are warnings and whose true values ​​are also warnings to all samples whose true values ​​are warnings.

[0078]

[0079] Among them, FN represents the number of samples that are actually warning but predicted as no warning, and the calculation method of F1 value is as follows.

[0080]

[0081] Step 3: Classification and Identification of Early Warnings

[0082] The hierarchical identification and early warning is to divide the levels according to the distribution of the current risk probability and the future risk probability. Specifically, the fusion early warning model is used to calculate the risk probability of disasters such as gas, fire, coal dust, roof, water, rock burst, coal and gas outburst at the current time point and in the future time period. If the disaster risk probability does not exceed the threshold in both time periods, no early warning will be issued. If the disaster risk probability at the current time point exceeds the threshold, and the disaster risk warning probability in the future time period also exceeds the threshold, a first-level early warning will be issued for the corresponding disaster, and a second-level early warning will be issued for other situations.

[0083] The present invention first collects a large amount of safety monitoring time series data from different coal mines as the initial training data set. Secondly, train Anomaly-Transformer and E 2 The GAN model identifies and interpolates abnormal data such as outliers and missing values ​​in time series data to achieve data preprocessing. Subsequently, a new multi-indicator fusion prediction model is introduced by combining the bidirectional long short-term memory network and the MMOE multi-task learning network. The preprocessed indicator data is input into the fusion prediction model after training and optimization, and the future development trend of each indicator can be obtained synchronously. Next, the multi-disaster risk fusion warning model based on the multi-task learning network is realized by improving the AdaTT model, and the number of model training iterations and the optimal risk warning threshold of each disaster are obtained when the F1 value is optimal by combining the F1 evaluation index. The real-time data and fusion prediction data of each monitoring indicator are input into the fusion warning model after training and optimization using field data, and the risk probability of each disaster at the current time point and the future time period can be obtained synchronously. Finally, a multi-disaster graded identification and warning method is proposed by combining the optimal risk warning threshold and the risk probability obtained by the fusion warning model. By comprehensively evaluating the risk warning status of different disasters in the current and future time periods, the risk graded identification and warning of each disaster is realized. The present invention constructs an intelligent fusion classification and grading advance warning method for major coal mine disasters, which mainly includes data preprocessing, multi-indicator fusion prediction, multi-disaster fusion warning and graded identification. This method can effectively respond to disaster risks such as gas, fire, dust, roof, water, rock burst, coal and gas outburst in a graded manner, and can provide risk information in advance, which is of great significance for accurate early warning of multi-hazard risks in coal mines.

Claims

1. A method for intelligent fusion classification and grading of coal mine major disasters, characterized in that The following steps are involved: Step 1: Data preprocessing; Collect a large amount of safety monitoring time series data in coal mines, including time series data of gas concentration, carbon monoxide concentration, coal dust concentration, fully mechanized mining resistance, surrounding rock separation, anchor cable stress, temperature, wind speed, water inflow, etc. as the initial data set, and use deep learning models to identify and interpolate abnormal data in the time series data to achieve data preprocessing; Step 2: Fusion prediction; A multi-indicator fusion prediction model is established based on a bidirectional long short-term memory network combined with an improved MMoE algorithm. The preprocessed data is input into the multi-indicator fusion prediction model. By learning the correlation between the indicators, the future development trends of different indicators are obtained simultaneously. Step 3: Fusion warning; Establish a multi-disaster risk fusion warning model based on the improved AdaTT multi-indicator fusion prediction model, input the original data and predicted data into the fusion warning model, use the precursor characteristics of historical and future data and the fusion analysis and identification of the relationship between disasters to simultaneously output the current and future risk probabilities of each disaster; The optimal model is determined through evaluation indicators, and the warning thresholds of various disaster risks are obtained based on the accuracy distribution in the test set; Step 4: Classification and identification; The current and future risk probabilities of various disasters obtained by the fusion warning model are combined with the risk warning thresholds of various disasters to establish a disaster risk classification identification and warning method that integrates current and future risk conditions. This method is used to carry out comprehensive risk identification and warning of disasters such as gas, fire, dust, roof, water, rock burst, coal and gas outbursts.

2. According to claim 1, a method for intelligent fusion classification and grading advance warning of major coal mine disasters is characterized in that: In step 1, the collected abnormal data is divided into training set, validation set and test set, and Anomaly-Transformer and E 2 GAN establishes a data pre-training model and performs training and testing. 2 GAN identifies and interpolates outliers and missing values ​​in the time series data of gas concentration, carbon monoxide concentration, coal dust concentration, comprehensive mining resistance, surrounding rock delamination, anchor cable stress, temperature, wind speed, water inflow, etc. collected on site.

3. The method for intelligent fusion classification and grading advance warning of major coal mine disasters according to claim 1 is characterized in that: In step 2, the method of establishing a multi-index fusion prediction model includes: Assume that all the input index data is x, and the index prediction task is K = {k1, k2, ...k n }, the number of expert networks is N, and the long-term dependency features of each indicator are extracted through a bidirectional long short-term memory network. The calculation method is as follows: in, is the candidate information memory cell of index k at time t, is the information memory cell of index k at time t, is the information output of the k index at time t; then the long-term dependency features of different indicators obtained by bidirectional long short-term memory network learning are spliced ​​to obtain the splicing result X c : X c =[H1,H2,…,H k ]; Among them, H k is the result of the kth indicator output by the bidirectional long-short time network; Finally, the spliced ​​matrix is ​​input into the multi-task learning model MMoE for fusion calculation, and the prediction results of each indicator are output synchronously after the task tower; the method is as follows: Among them, G k (x) is the gate control unit of task k, W gk To learn the parameter weights, f k (x) is the output of task k after learning through the parameter-sharing expert network, m k For the mission tower network, y k It is the prediction result of k tasks obtained by fusion analysis of all indicator data.

4. The method for intelligent fusion classification and grading advance warning of major coal mine disasters according to claim 1 is characterized in that: In step three, the method for determining the warning threshold of each disaster includes: Assume that the input data of each disaster risk monitoring indicator is x, and its warning task is T = {t1, t2, ..., t n }, the number of fusion unit stacking layers is L = {l1, l2, ..., l n }, the t-th task fusion module in the l-th layer is represented as Then the fusion output of the t-th warning task at the l-th layer is for: The calculation process of each fusion network block FU is to first integrate the outputs of multiple expert networks of key indicators through a linear model. The formula is as follows: in, After superimposing the matrix for each expert, is the parameter vector, NELF is the output of multiple experts of key indicators after linear fusion; then, the gating network is used to learn the correlation between other indicators and key indicators, and all the expert network outputs are fused to obtain the influence weight of each factor on the evolution of disaster risk. The formula is as follows: Finally, the gating network results are added to the results of the key indicator expert network fusion to obtain a data fusion block for the disaster risk warning task. The calculation formula is as follows:

5. The method for intelligent fusion classification and grading advance warning of major coal mine disasters according to claim 1 or 4, characterized in that: In step three, the optimal model of the warning model is determined by the optimal F1 value method. The model determination method based on the threshold of F1 value maximization is used to obtain the number of model iterations when F1 is optimal in the iterative training process and determine its corresponding threshold in combination with the accuracy distribution of the optimal model in the test set. F1 is the harmonic mean of precision and recall. The values ​​of precision and recall are further modified by adjusting the threshold so that the F1 value reaches the maximum, thereby determining the warning thresholds for each disaster risk.

6. The method for intelligent fusion classification and grading advance warning of major coal mine disasters according to claim 5 is characterized in that: The F1 value is determined as follows: Suppose the precision is P. The precision P is the ratio of the number of results that are actually marked as warnings among the results that are output as warnings by the model to the number of results that are output as warnings by the model. The calculation is as follows: Among them, TP represents the number of samples that are actually warning and predicted as warning, and FP represents the number of samples that are actually not warning but predicted as warning; Suppose the recall rate is R, which refers to the proportion of samples whose predicted value is warning and whose true value is also warning to all samples whose true value is warning: Among them, FN represents the number of samples that are actually warning but predicted as no warning, and the calculation method of F1 value is as follows:

7. The method for intelligent fusion classification and grading advance warning of major coal mine disasters according to claim 1 is characterized in that: In step four, the fusion warning model is used to calculate the precursor characteristics of historical and future data to obtain the risk probabilities of disasters such as gas, fire, coal dust, roof, water, rock burst, coal and gas outbursts at the current time point and in the future time period. If the disaster risk probability does not exceed the threshold in both time periods, no warning will be issued. If the disaster risk probability at the current time point exceeds the threshold and the disaster risk warning probability in the future time period also exceeds the threshold, a first-level warning will be issued for the corresponding disaster, and a second-level warning will be issued for other situations.

Images