Multi-element fusion coal mine disaster risk intelligent identification and early warning method and system

Abstract

The invention discloses a multivariate fusion coal mine disaster risk intelligent identification and early warning method and system, and belongs to the technical field of coal mine major disaster risk intelligent identification and early warning. The method comprises the following steps: collecting multi-index data of a coal mine working face as an initial data set, and preprocessing time series data by using a one-dimensional depth channel attention residual convolution-LSTM model and a chain-based multi-interpolation model; performing Pearson's correlation coefficient and Spearman level correlation analysis on the preprocessed data, and determining different major disaster risk monitoring and early warning index sets; risk precursor, interference and noise signals of different indexes are analyzed and marked, a sample enhancement model based on the de-noising diffusion probability is input to complete sample expansion, an MACN model is trained, and a multivariate index risk fusion identification method based on the MAKER and a comprehensive early warning rule are established for major disasters. The method has the advantages of dynamic adaptability, configurability, high interpretability and reliability under the condition of small samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent identification and early warning of major disaster risks in coal mines, and in particular to a multi-element fusion intelligent identification and early warning method and system for disaster risks in coal mines. BACKGROUND

[0002] The prior art analyzes the precursor characteristics of each disaster monitoring signal under different environments, and uses a single or multiple indicators to independently identify each major disaster risk through statistical methods such as trend method, critical value method, and multi-condition judgment method. However, with the increase in mining depth and the increasing complexity of geological conditions in mines, the coupling of multiple risks is serious. This type of early warning cannot accurately quantify the evolution response characteristics of each disaster, and is prone to missed reports and false reports. Its generalization and reliability are relatively low. At the same time, in the prior art, deep learning models such as CNN, RNN, Transformer, or improved models are used for early warning. The use of these models often results in insufficient training samples or unbalanced sample sets due to the lack of dangerous data samples, affecting the performance of model training and the accuracy of early warning. More importantly, there are interference, noise, and other signals in the field data. Directly inputting these signals into the above deep learning models may cause more risk false positives and result in a large amount of computational cost. In addition, the deep learning model for early warning is difficult to integrate with artificial evaluation and dynamic configuration. The parameters of the model are fixed after training, and the deep coal mine environment is complex. Disasters have complex causes and various inducing factors. This type of early warning method has poor dynamic adaptability and is not highly interpretable. SUMMARY

[0003] In view of the deficiencies of the prior art, a multi-element fusion intelligent identification and early warning method and system for disaster risks in coal mines are provided. The method realizes intelligent identification and early warning of major disaster risks in coal mines based on deep learning under small samples.

[0004] To achieve the above technical problems, the technical solution provided by the present application is as follows: a multi-element data fusion-based intelligent identification and early warning method for major disaster risks in coal mines, comprising the following steps: S1, collect a large number of real samples of disasters in coal mine working faces to construct a time series data set, use a one-dimensional deep channel attention residual convolution-LSTM model based on self-supervised learning to identify abnormal data in the time series data set, use a multiple imputation model MICE to impute missing data in the time series data set, and form a real sequence composed of index non-abnormal value data; S2, perform Pearson correlation coefficient and Spearman rank correlation analysis on the preprocessed time series data set to determine different coal mine major disaster risk monitoring and early warning indicators; S3, analyze the coal mine major disaster risk monitoring and early warning index, input the risk precursors, interference and noise signals obtained by analyzing the various coal mine working face before the risk into the sample enhancement model DDPM based on denoising diffusion probability, expand the samples of different early warning indexes and their corresponding coal mine major disaster risks; S4, train the one-dimensional multi-scale multiple attention parallel time deep convolutional neural network MACN using the expanded samples of different early warning indexes and their corresponding coal mine major disaster risks, identify the precursor signal recognition result of the coal mine major disaster risk, and distinguish and exclude interference and noise signals; S5, use the maximum likelihood method evidence reasoning MAKER to construct a multi-index risk fusion identification model of various major disasters, and use the precursor signal recognition result to complete the training of the multi-index risk precursor identification model of various major disasters; S6, use the multi-index risk precursor identification model of major disasters to establish a major disaster risk intelligent identification and early warning system, realize intelligent application of coal mines and real-time monitoring and early warning of major disaster risks.

[0005] Further, the real samples of the coal mine working face disaster include multi-index data: gas concentration data, carbon monoxide concentration data, coal dust concentration data, fully mechanized resistance data, surrounding rock separation data, anchor rod stress data, temperature data, wind speed data, electromagnetic radiation data, acoustic emission data, long observation hole and water flow data.

[0006] Further, the one-dimensional deep channel attention residual convolution-LSTM model based on self-supervised learning includes an encoder and a decoder, wherein the output of the one-dimensional deep channel attention residual convolution in the encoder is connected to the input of the LSTM model, and the output of the LSTM model in the decoder is connected to the input of the one-dimensional deep channel attention residual convolution; The real sequence composed of index data without abnormal values is input into the encoder to generate an intermediate feature matrix, then the feature matrix is input into the decoder, the decoder reconstructs the sequence according to the intermediate features, and repeatedly performs self-supervised learning training until the reconstructed sequence without abnormal values is reconstructed, and the abnormal values can be identified through the determination threshold; the one-dimensional deep channel attention residual convolution is based on the time domain channel attention and frequency domain three-part energy cross-gate fusion mechanism of one-dimensional residual convolution, realizes high-sensitive modeling of sequence spectrum shape change, and combines LSTM to form a self-supervised structure, realizes multi-scale time-frequency joint modeling of time sequence characteristics of working face indexes and accurate identification of abnormal points; Let the one-dimensional time sequence constructed by the coal mine major disaster risk monitoring and early warning index as a real sequence input , wherein is the batch size, T is the time step of the input sequence, and the original sequence can be upgraded to C channels using the 1D convolution operation formula: , in Let K be the weight of the c-th output channel relative to the input data at position j of the convolution kernel, and K be the kernel size. For bias units, output dimension C is the number of output channels. The convolution uses the same padding to keep the length of the output sequence unchanged. t is the time step position of the output after convolution. The convolutional backbone first uses two layers of convolution and non-linearity to extract feature information from the matrix Y output by the 1D convolution, which contains different receptive fields. , in Let the kernel be a one-dimensional convolution kernel of length k and dilation rate d. , For two one-dimensional convolutional kernels of different sizes, channel attention is calculated on the main branches of the convolution, including channel compression and frequency domain cross-gating; first, the feature information is... Perform channel compression: , , , in This represents the mean of the c-th channel over the time dimension. , It is a non-linear activation function. This represents the scaling factor for the c-th channel, generated by a fully connected + Sigmoid algorithm. , These represent the weight matrices of the fully connected layer, This is the feature vector of the c-th channel output by the main branch of the convolution; Perform a Fast Fourier Transform (rFFT) on the input data for each channel and take the amplitude: , , in Let f(x) be the FFT coefficient of the x-th channel at the x-th frequency. For Fourier transform, The amplitude is then used as the basis for determining the frequency index axis of each channel after the FFT, which is then divided into three index sets. Define frequency band energy: , in The average energy vectors of the low, mid, and high frequency bands of the nth channel are obtained; then, the gating weights are obtained by sharing the average energy vectors of the low, mid, and high frequency bands using a multilayer perceptron. , in , This is the weight matrix. For bias terms, The frequency domain gating factor generated for the nth channel; Finally, the gating weights are merged with the channel compression results: , The feature vectors of the coal mine major disaster risk monitoring and early warning indicators after dimensional matching are added to the fusion result to form a residual connection: , This is the feature vector after dimensional matching processing of the monitoring and early warning indicators for major coal mine disaster risks. The output of the first-layer depth channel attention residual convolution module is used as the input of the next layer; the number of layers needs to be set according to the complexity of different index fluctuations, and can be set between 3 and 8 layers. The encoder finally outputs a one-dimensional depth channel attention residual convolution with the dimension of... The data is input into the LSTM; the LSTM output is... The intermediate feature matrix is ​​generated; the input of the one-dimensional deep channel attention residual convolution of the decoder is the output of the LSTM. ; The loss L1 is calculated based on the reconstructed sequence values ​​from intermediate features and the true sequence values ​​from the real, anomaly-free data. L1 is then used to evaluate the reconstructed sequence values ​​and reconstruction error generated by the LSTM model. , in The reconstructed sequence values ​​output by the decoder. This is the actual sequence value input to the encoder.

[0007] Furthermore, the workflow of the chain-based multiple interpolation model for MICE is as follows: Let the data variables in the time series data be represented as The MICE model first fills the missing values ​​of each variable with random values. In each subsequent iteration, it uses other variables in the time series to fill the missing values ​​of the current variable. Other variables include: gas concentration data, carbon monoxide concentration data, coal dust concentration data, fully mechanized mining resistance data, surrounding rock delamination data, anchor cable stress data, temperature data, wind speed data, electromagnetic radiation data, acoustic emission data, long borehole and water flow rate. For variables The The interpolated values ​​for the next iteration are obtained through the following distribution: , where MICE model parameters at the i-th iteration the complete conditional posterior distribution of is the prior distribution of the posterior predictive distribution is calculated using the i-th iteration the updated imputed value: , where is the probability of observing , under the MICE model for the variable , denotes the observed non-missing sample value, denotes the values of all other variables except the current variable where is the imputed or updated value, is the unimputed or updated value.

[0008] Further, by Pearson correlation coefficient, Spearman rank correlation coefficient on the collection of coal mine major disaster data before the period of analysis, get gas disaster and gas concentration, coal dust concentration, fully mechanized resistance, surrounding rock separation, anchor rod stress, temperature, wind speed, electromagnetic radiation and acoustic emission index have correlation; fire and gas concentration, carbon monoxide concentration, temperature and wind speed are related, dust and water disaster long observation hole and water flow are related; roof and other coal and rock dynamic disaster and gas concentration, coal dust concentration, fully mechanized resistance, surrounding rock separation, anchor rod stress, electromagnetic radiation and acoustic emission are related; based on this, the index set of different major disasters is established.

[0009] Further, the one-dimensional multi-scale multi-attention parallel time sequence deep convolutional neural network MACN uses multi-scale convolution to capture features of different time scales, and forms branch attention, channel attention and time attention based on the lightweight weight generation attention to fine weighting of features, and finally uses a residual network to reduce the information loss of multi-layer superposition: Downsample and feature dimension of data input into MACN network: the input data executes convolutional neural network, batch normalization layer and ReLU activation function in turn, and then uses the maximum pooling layer to output the pre-processing result; The pre-processing result is input into the feature extraction module composed of multiple multi-scale convolution attention blocks. The core process of each multi-scale convolution attention block is: dimension mapping of one-dimensional convolution is performed on the input data, then a branch is established for each convolution kernel, and each branch outputs a group of branch vectors where For dimension mapping output, For output channel size, Each branch of the calculation network, the calculation network includes one-dimensional convolution, batch normalization and activation function; Stack all branch data into data tensor Wherein The number of branches; The tensor input branch attention, dynamically selects the contribution of different receptive fields; branch attention first averages the time dimension, then obtains branch attention weight through multi-layer perception MLP and softmax , then the weight of each branch is combined, the expression is: , , Wherein Branch attention result, The slice vector of the bth branch, the branch attention result is input into the channel attention, which amplifies the channel with information quantity for the task and suppresses irrelevant channels; the channel attention first obtains the feature description of each channel through global time sequence pooling, then uses linear layer-activation function-linear layer and softmax to obtain channel attention weight , finally, the recalibration is performed according to the channel attention weight, and the calculation formula is as follows: , , Wherein The result of channel j through channel weight Hadamard product, the obtained data tensor is input into the time domain attention, the time domain attention splices the features of all branches in the channel to form a data tensor , then the time sequence attention weight of each time point is realized through one-dimensional convolution combined with softmax function , then the time sequence weighting of each channel is realized: , , Finally, the one-dimensional convolution and batch normalization projection output channel dimension, and then add the residual, the calculation method is as follows: , If it is not the last layer, it is used as the input of the next layer, otherwise, the classification result is output through global average pooling and classification layer: , The risk, interference and noise signal types of the expanded index data set are marked, and input into the MACN model for supervised precursor signal multi-classification identification model training. The MACN model trained according to the real-time monitoring signal changes can effectively classify the dangerous, interference and noise signals, and filter out the dangerous signals and exclude other interference and noise signals.

[0010] Further, the multi-index risk fusion identification model of various major disasters identifies the precursor risks of coal mine major disasters according to various sensor information, processes the uncertainty through basic probability distribution, belief function and likelihood function by using state space model and evidence space model, and fuses the multi-source sensor data by using Dempster combination rule; The output result of the multi-index risk fusion identification model of various major disasters is represented as the precursor risk probability information of each major disaster, so that the discrimination result output by the multi-index risk fusion identification model of various major disasters is converted into basic probability distribution, and the support degree for each precursor risk state subset is quantified as: , wherein, represents the basic probability distribution of the i-th feature supporting a specific state, and represent the basic probability of the risk and non-risk states in the state subset; In order to measure the mutual dependence between different sensor information sources, a joint dependence index is introduced: , wherein is the joint likelihood approximation, and if it reflects the conflict; is the joint dependence coefficient. The higher the dependence coefficient is, the stronger the consistency of the two information sources on the focal element G and S is. G and S represent the state subset supported by the evidence and to be identified, including {risk}, {non-risk} or their combination, is the joint likelihood measure of the i-th sensor information source and the j-th sensor information source, , respectively represent the likelihood of the sensor information source i to the focal element G and the sensor information source j to the focal element S. Suppose that three sensor information sources are A, B and E. Under the condition that the basic probability distribution and the mutual dependence coefficient are known, the joint support degree of the three sensor information sources is calculated by using the following formula: , wherein Let F be the joint dependency index of the three sensors, and F be the fused focal element, power set. for{ {Risk}, {No Risk}, {Risk, No Risk}}; the fusion contribution is only valid if the intersection of the three is not empty; for all combinations with empty intersections, their quality is considered conflict quality: , Finally, the mass of all non-empty focal elements is normalized using the following formula: where the focal element is the aforementioned power set { A subset of {risk}, {no risk}, {risk, no risk}; , Subset-level reliability weights for the three sensors The training process aims to minimize the error between the fusion result and the true label. , in This is the s-th sample, including sensor data collected by sensors A, B, and E, and the fused basic probability assignment of the corresponding real disaster state labels. The label is the real label. The minimization problem is solved using the finite-memory quasi-Newton method L-BFGS-B. Finally, in the fused basic probability assignment, the single-element focal element with the highest support is selected as the predicted label. , If the maximum support falls on the power set {risk, no risk}, it is considered a "fuzzy / uncertain" state, and the decision remains undetermined; This function outputs whether there is a risk or uncertainty regarding the current disaster.

[0011] Furthermore, establish comprehensive disaster early warning rules to better achieve configurable and dynamic comprehensive early warning in complex coal mine conditions: 1) If no signal in the disaster indicator set is identified as indicating a precursor risk, it is considered that there is no risk of outbreak, and no risk will be marked, and no warning will be issued; 2) If, among the key indicators specified by experts, a single sensor information source is identified as having a precursor risk, then the current time will be marked as a potential precursor to a disaster risk, and a Level 1 warning will be issued; 3) If the identified precursory risk persists for more than 15 minutes, or if two or more sensor information sources with precursory characteristics of disaster risk are detected simultaneously among the key indicators specified by experts, or if the presence of risk is output according to the MAKER discrimination method, then the situation is marked as a level 2 warning.

[0012] A risk intelligent identification and early warning system of a coal mine major disaster risk intelligent identification and early warning method based on multi-element data fusion, comprising a coal mine data access unit, a monitoring data visualization unit, an early warning information visualization unit, a model online incremental training unit, a model interface configuration unit and an early warning rule configuration unit; The coal mine data access unit accesses the data of the coal mine monitoring sensor by configuring relevant parameters, queries the current and historical data of all accessed sensors in the coal mine in real time, supports filtering sensors of different locations and types, supports querying and displaying historical data according to different time ranges, and selects relevant monitoring indicators and their data to perform online incremental training on the model; The monitoring data visualization unit is used to display the sensor information source through the industrial computer display; The early warning information visualization unit is used to record the history of the multi-element data fusion coal mine major disaster risk intelligent identification and early warning method during the on-site application process, and filter and query the historical early warning data by specifying the time range and the early warning level; The model online incremental training unit is used to train and adjust the diffusion sample generation model, the MACN precursor risk identification model and the MAKER method in the multi-element data fusion coal mine major disaster risk intelligent identification and early warning method, automatically record the last training data node, or manually set the time range, select the specified model for training, automatically save the historical iteration version after training is completed, and transmit the result to the data platform, and save the training information to the database; The model interface configuration unit is used to configure the model interface parameters, permission control and model selection switching information, and call the interfaces of diffusion model sample generation, precursor risk identification and multi-element fusion early warning; The early warning rule configuration unit can manually configure the risk early warning rules according to different situations on site, including the definition rules of no early warning, first-level early warning and second-level early warning in the early warning rules.

[0013] Beneficial effects: the one-dimensional deep channel attention residual convolution-LSTM model proposed in the application, wherein the one-dimensional deep channel attention residual convolution is based on a time domain channel attention and a frequency domain three-part energy cross-gate fusion mechanism of a one-dimensional residual convolution, realizes high-sensitive modeling of sequence spectrum shape changes, and forms a self-supervised structure by combining the LSTM to realize multi-scale time-frequency joint modeling of time sequence characteristics and accurate identification of abnormal points. The application uses a one-dimensional diffusion model to enhance the diversity of dangerous data samples, ensuring sufficient samples to train the deep learning model. Again, a one-dimensional multi-scale multi-attention parallel time sequence deep convolutional neural network MACN is proposed to combine maximum likelihood evidence reasoning and rules to realize risk early warning, wherein the MACN identifies precursor characteristics before early warning, can exclude the influence of interference and noise signals on subsequent early warning, reduces the calculation cost of subsequent early warning models or methods, in addition, the MAKER and rule early warning can fuse multi-sensor / multi-source data for early warning, and has the ability to handle uncertainty and incomplete information, and can integrate expert knowledge, rules and subjective judgment rules. The application preprocesses the collected index data to provide high-quality sequence data for subsequent data enhancement and training of precursor identification models, and then expands precursor, interference, noise and other samples of different indicators, providing a large amount of data basis for precursor identification deep learning model training and fusion early warning and rule making. Then use the MACN model to train and identify the precursor characteristics of each indicator, improve the risk early identification capability, and can identify and exclude interference, noise and other signals, reduce the influence on early warning accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is the process block diagram of the multi-element fusion coal mine disaster risk intelligent identification and early warning method in the embodiment of the application.

[0015] Figure 2 is a one-dimensional deep channel attention residual convolution-LSTM model structure diagram in the embodiment of the application.

[0016] Figure 3 is a one-dimensional deep channel attention residual convolution-LSTM model flow diagram in the embodiment of the application.

[0017] Figure 4 is a MACN model execution flow diagram in the embodiment of the application. DETAILED DESCRIPTION

[0018] The embodiments of the application will be further described below in combination with the drawings: As Figure 1 shown, the application discloses a multi-element fusion coal mine disaster risk intelligent identification and early warning method, the steps are as follows: S1, a large number of real samples of coal mine working face disasters are collected to construct a time series dataset, a one-dimensional deep channel attention residual convolution-LSTM model based on self-supervised learning is used to identify abnormal data in the time series dataset, a multiple imputation model MICE is used to impute missing data in the time series dataset, and a real sequence composed of index non-abnormal value data is formed; The real samples of the coal mine working face disasters include multi-index data: gas concentration data, carbon monoxide concentration data, coal dust concentration data, fully mechanized resistance data, surrounding rock separation data, anchor rod and cable stress data, temperature data, wind speed data, electromagnetic radiation data, acoustic emission data, long observation hole and water flow data.

[0019] As shown in Figure 2 and Figure 3 , the one-dimensional deep channel attention residual convolution-LSTM model based on self-supervised learning includes an encoder and a decoder, wherein the output of the one-dimensional deep channel attention residual convolution in the encoder is connected to the input of the LSTM model, and the output of the LSTM model in the decoder is connected to the input of the one-dimensional deep channel attention residual convolution; The encoder inputs the real sequence composed of index non-abnormal value data, first generates an intermediate feature matrix through the encoder, then inputs the feature matrix into the decoder, the decoder reconstructs the sequence according to the intermediate feature, repeatedly performs self-supervised learning training, until a reconstructed sequence without abnormal values is reconstructed, and the abnormal values can be identified through a judgment threshold; the one-dimensional deep channel attention residual convolution is based on the time domain channel attention and frequency domain three-part energy cross-gate fusion mechanism of one-dimensional residual convolution, realizes high-sensitive modeling of sequence spectrum shape change, combines with LSTM to form a self-supervised structure, and realizes multi-scale time-frequency joint modeling of time series characteristics of working face indexes and accurate identification of abnormal points; Set the one-dimensional time series constructed as a real sequence input , wherein is the batch size, T is the time step of the input sequence, and the original sequence can be upgraded to C channel number using the 1D convolution operation formula: , , wherein is the weight of the input data at the convolution kernel position j for the cth output channel, K is the convolution kernel size, is the bias unit, the output dimension , C is the number of output channels, the convolution uses the same padding to keep the output sequence length unchanged, and t is the time step position of the output after convolution. In the convolution main branch, two layers of convolution and nonlinearity are first used to extract feature information of different receptive fields for the 1D convolution output matrix Y: , , wherein is a one-dimensional convolution kernel with length k and expansion rate d, , is a one-dimensional convolution kernel with two different sizes, and the channel attention calculation is performed on the convolution stem branch, including channel compression and frequency domain cross-gate composition; first, the feature information is compressed in the channel: , , , wherein represents the mean value of the cth channel in the time dimension, , is a nonlinear activation function, represents the scaling factor of the cth channel, which is generated by a full connection + Sigmoid; , respectively represent the full connection layer weight matrix, is the feature vector of the cth channel output by the convolution stem branch; Fast Fourier Transform rFFT is performed on the data input to each channel, and the amplitude is taken: , , wherein is the FFT coefficient of the cth channel at the kth frequency, is the Fourier transform, is the amplitude, and then the frequency index axis of each channel after FFT is divided into three index sets , and the frequency band energy is defined: , wherein is the low, medium and high three-band energy mean vector of the cth channel; then the shared multilayer perceptron is performed on the low, medium and high three-band energy mean vector to obtain the gating weight: , wherein , is the weight matrix, is the bias term, is the frequency domain gating factor generated by the cth channel; Finally, the gating weight is fused with the channel compression result: , The residual connection is formed by the fusion result and the feature vector after the dimension matching processing of the coal mine major disaster risk monitoring and early warning index after the dimension matching processing. , is the feature vector after dimension matching processing of the monitoring and early warning index of the major disaster risk of the coal mine, is the output of the first layer of the deep channel attention residual convolution module and serves as the input of the next layer; the number of layers is set according to the fluctuation complexity of different indexes, and is set to 3-8 layers; The dimension of the one-dimensional deep channel attention residual convolution output of the encoder is ; the LSTM output is , generating an intermediate feature matrix; the input of the one-dimensional deep channel attention residual convolution of the decoder is the output of the LSTM ; wherein the main principle formula of the LSTM core module mainly includes an input gate, a forgetting gate, an output gate, a candidate memory, a current memory state, and a current hidden state, and the main principle formula is as follows: , wherein , are the input dimension and the output dimension respectively, the LSTM output is , generating an intermediate feature matrix, and then the input of the one-dimensional deep channel attention residual convolution of the decoder is the output of the LSTM .

[0020] According to the intermediate feature reconstruction sequence value and the real sequence value of the real non-abnormal data, the loss calculation result L1 is calculated, and the reconstruction sequence value of the LSTM model is used to judge the reconstruction error: , wherein is the reconstruction sequence value output by the decoder, is the real sequence value input into the encoder.

[0021] Workflow: input the non-abnormal value time sequence, judge whether the length reaches the preset threshold, if not, use the regression model to upgrade to the threshold, which is 512 in this scheme; if it reaches, use the current sequence length; add the sine position coding to the sequence, use the one-dimensional residual convolution-LSTM encoder to process the intermediate coding, and then use the one-dimensional residual convolution-LSTM reconstructor (decoder) to reconstruct the sequence; compare the reconstruction sequence with the original sequence to calculate the error, judge whether the error is lower than the threshold or stable, determine the error threshold; input the new data point into the database, judge whether the error is higher than the threshold, according to the threshold to judge whether the current data point is an abnormal point sequence or a non-abnormal value sequence, if it is, delete the abnormal point to form a missing value.

[0022] The working process of the MICE model is as follows: Let the data variables in the time series data be represented as The MICE model first fills the missing values ​​of each variable with random values. In each subsequent iteration, it uses other variables in the time series to fill the missing values ​​of the current variable. Other variables include: gas concentration data, carbon monoxide concentration data, coal dust concentration data, fully mechanized mining resistance data, surrounding rock delamination data, anchor cable stress data, temperature data, wind speed data, electromagnetic radiation data, acoustic emission data, long borehole and water flow rate. For variables The The interpolated values ​​for the next iteration are obtained through the following distribution: , in MICE model parameters at the next iteration The fully conditional posterior distribution, for The prior distribution; Calculate using posterior prediction distribution iteration Updated interpolation values: , in For variables , Observed under the MICE model The probability, This represents the observed, non-missing sample values. This indicates the values ​​of all variables except the current variable. ,in It is an interpolated or updated value. It is an uninserted or unupdated value.

[0023] S2. Perform Pearson correlation coefficient and Spearman rank correlation analysis on the preprocessed time-series dataset to determine the monitoring and early warning indicators for major disaster risks in different coal mines. This invention uses Pearson correlation coefficient and Spearman rank correlation coefficient to analyze indicators related to different major disasters from different perspectives. The main theoretical methods are as follows: The Pearson correlation coefficient is primarily used to measure the linear relationship between two variables, that is, whether the other variable exhibits the same or opposite trend when one variable shows a linear change. It can determine the strength and direction of the linear relationship between the indicators. Therefore, the Pearson correlation coefficient ranges from [-1, 1], where 1 represents perfect positive correlation, -1 represents perfect negative correlation, and 0 represents no correlation. Assuming the two time-series variables are x and y, the calculation formula is as follows: , in , are the values of two variables at time point i, respectively, , are the mean values of two variables, and n is the data size of the variable set.

[0024] Spearman rank correlation coefficient can measure the monotonic relationship between two variables, which may not be linear. Compared with Pearson correlation coefficient, this method first changes the original data into corresponding ranks, and then calculates the Pearson correlation coefficient between these ranks. Assuming that two time series variable sets are x and y, the calculation formula is as follows: , wherein is the difference between the two variables after calculating the corresponding ranks at time point i. According to the above formula, is inversely proportional to , that is, when the difference between the two variables after calculating the ranks is very small, tends to 1, and on the contrary tends to -1, which respectively represent completely consistent ranking and completely reversed ranking.

[0025] Through the analysis of the collected data by Pearson correlation coefficient and Spearman rank correlation coefficient, it is found that gas disaster is related to gas concentration, coal dust concentration, fully mechanized mining resistance, surrounding rock separation, anchor rod stress, temperature, wind speed, electromagnetic radiation and acoustic emission indicators; fire is related to gas concentration, carbon monoxide concentration, temperature and wind speed; dust is related to long observation hole and water flow in water disaster; roof and other coal and rock dynamic disasters are related to gas concentration, coal dust concentration, fully mechanized mining resistance, surrounding rock separation, anchor rod stress, electromagnetic radiation and acoustic emission. Based on this, the index set of different major disasters is established.

[0026] S3, analyze the coal mine major disaster risk monitoring and early warning index, input the risk precursors, interference and noise signals obtained by analyzing the various coal mine working face before the risk into the sample enhancement model DDPM based on denoising diffusion probability, expand the samples of different early warning indexes and their corresponding coal mine major disaster risks; the sample enhancement model based on the diffusion probability model (DDPM) is composed of 1DUnet, self-attention mechanism and Gaussian diffusion. By inputting the original data sample, the model gradually down-samples the real data and adds Gaussian noise, and after several rounds of increase, it is finally completely noisy, close to pure Gaussian noise, and realizes the forward diffusion process. Then the model restores the data step by step through up-sampling of the completely noisy data, so that the generated data approximates the distribution characteristics of the original data. By using its learning sequence distribution characteristics of risk, interference and noise, a large number of sample sequences of different types with similar distribution to the original sequence can be generated. For the generated sequence, the sequence statistical characteristics are calculated by kurtosis, margin factor, slope, variance and standard deviation, and compared and analyzed with the original sequence, and the signals meeting the distribution characteristics are selected and added to the sample set. The principle of the statistical method is as follows: , , , , .

[0027] S4, train the one-dimensional multi-scale multi-attention parallel time series deep convolutional neural network MACN using the expanded samples of different early warning indexes and their corresponding coal mine major disaster risks, identify the precursor signal recognition result of the coal mine major disaster risk, and distinguish and exclude interference and noise signals.

[0028] As shown in Figure 4 , the one-dimensional multi-scale multi-attention parallel time series deep convolutional neural network MACN uses multi-scale convolution to capture features of different time scales, and uses branch attention, channel attention and time attention formed by light weight attention generation to finely weight the features, and finally uses a residual network to reduce the information loss of multiple layers of superposition: Down-sampling and feature dimensioning of data input into the MACN network: the input data are sequentially executed with a convolutional neural network, a batch normalization layer and a ReLU activation function, and then a max-pooling layer is used to output a pre-processing result; The pre-processing result is input into a feature extraction module composed of multiple multi-scale convolution attention blocks, and the core process of each multi-scale convolution attention block is: dimension mapping of the input data is performed by one-dimensional convolution, then a branch is established for each convolution kernel, and each branch outputs a group of branch vectors , wherein for dimension mapping output, for output channel size, a calculation network for each branch, the calculation network comprising a one-dimensional convolution, a batch normalization and an activation function; stacking all branch data into a data tensor wherein is the number of branches; inputting the tensor into branch attention to dynamically select the contribution of different receptive fields; the branch attention first averages the time dimension, and then obtains branch attention weights through a multi-layer perception (MLP) and a softmax , and then weights are combined for each branch, and the expression is: , , wherein is the branch attention result, which is input into channel attention to amplify channels with information quantity for the task and suppress irrelevant channels; the channel attention first obtains feature descriptions of each channel through global time pooling, and then obtains channel weights through a linear layer-activation function-linear layer and a softmax , and finally recalibrates according to the channel weights. is the slice vector of the bth branch, the branch attention result is input into channel attention to amplify channels with information quantity for the task and suppress irrelevant channels; the channel attention first obtains feature descriptions of each channel through global time pooling, and then obtains channel attention weights through a linear layer-activation function-linear layer and a softmax , and finally recalibrates according to the channel attention weights, and the calculation formula is as follows: , , wherein is the result of channel j through channel weight Hadamard product, and the obtained data tensor is input into time domain attention; the attention can make the model pay attention to key time steps (such as key classification features, local area mutations, etc.), and meanwhile, the attention can realize cross-scale information interaction and then gather and focus time, so as to better fuse multi-element data. The time domain attention splices the features of all branches in the channel to obtain a data tensor , and then realizes time sequence attention weights of each time point through one-dimensional convolution combined with a softmax function , and then realizes time sequence weighting for each channel: , , Finally, the output channel dimension is projected by one-dimensional convolution and batch normalization, and then added with the residual, which is calculated as follows: , If it is not the last layer, it is the input of the next layer, otherwise, the classification result is output by global average pooling and classification layer: , The expanded index data set is marked with risk, interference and noise signal types, and input into the MACN model for supervised precursor signal multi-classification recognition model training. The trained MACN model can effectively classify the dangerous, interference and noise signals according to the real-time monitoring signal changes, and filter out the dangerous signals and exclude other interference and noise signals.

[0029] S5, using the maximum likelihood method evidence reasoning MAKER to build a multi-index risk fusion identification model of various major disasters, and using the precursor signal recognition result to complete the training of the multi-index risk precursor identification model of various major disasters; Using the multi-index risk fusion identification model of various major disasters, the precursor risk of coal mine major disasters is identified according to the information of each sensor. The state space model and the evidence space model are used to process uncertainty through basic probability assignment, belief function and likelihood function. Dempster combination rule is used for multi-source sensor data fusion; The output result of the multi-index risk fusion identification model of various major disasters is represented as the precursor risk probability information of each major disaster, so the discrimination result output by the multi-index risk fusion identification model of various major disasters is converted into basic probability assignment, and the support degree of each precursor risk state subset is quantified as: , In the formula, Pi represents the basic probability distribution of the i-th feature supporting a specific state, and represent the basic probability of risk and non-risk state in the state subset; In order to measure the mutual dependence between different sensor information sources, the joint dependence index is introduced: , Where is the joint likelihood approximation, and if it reflects the conflict; is the joint dependence coefficient. The higher the dependence coefficient, the stronger the consistency of the two information sources on the focal element G, S. G, S represents the state subset supported by the evidence and to be identified, including {risk}, {non-risk} or their combination, Let be the joint likelihood measure of the information source i and the information source j. , Let G and S represent the likelihood of focusing elements G and S, respectively, for sensor information source i and sensor information source j. Let A, B, and E be the three sensor information sources. Given the basic probability distribution and interdependence coefficients... Under the given conditions, the joint support of the three sensor information sources is calculated using the following formula: , in Let F be the joint dependency index of the three sensors, and F be the fused focal element, power set. for{ {Risk}, {No Risk}, {Risk, No Risk}}; the fusion contribution is only valid if the intersection of the three is not empty; for all combinations with empty intersections, their quality is considered conflict quality: , Finally, the mass of all non-empty focal elements is normalized using the following formula: where the focal element is the aforementioned power set { A subset of {risk}, {no risk}, {risk, no risk}; , Subset-level reliability weights for the three sensors The training process aims to minimize the error between the fusion result and the true label. , in This is the s-th sample, including sensor data collected by sensors A, B, and E, and the fused basic probability assignment of the corresponding real disaster state labels. The label is the real label. The minimization problem is solved using the finite-memory quasi-Newton method L-BFGS-B. Finally, in the fused basic probability assignment, the single-element focal element with the highest support is selected as the predicted label. , If the maximum support falls on the power set {risk, no risk}, it is considered a "fuzzy / uncertain" state, and the decision remains undetermined; This function outputs whether there is a risk or uncertainty regarding the current disaster.

[0030] S6. Establish a major disaster risk intelligent identification and early warning system using a multi-indicator risk precursor identification model for major disasters, realizing intelligent applications in coal mines and real-time monitoring and early warning of major disaster risks: Set up an intelligent detection and early warning system to better achieve configurable and dynamic comprehensive early warning in complex coal mine conditions. 1) If no signal is identified as showing a precursor risk in the index set of disasters, it is considered that there is no outbreak risk, and no risk is marked, and no warning is output; 2) If a single sensor information source is identified as having a precursor risk in the expert-specified index with key discriminant, the current time will be marked as having a potential disaster risk precursor, and a first-level warning will be given; 3) If the identified precursor risk lasts more than 15 minutes, or at the same time, two or more sensor information sources in the expert-specified index with key discriminant are detected to have disaster risk precursor characteristics, or the existence of risk is output according to the MAKER discriminant method, this case is marked as a second-level warning.

[0031] The system of the coal mine major disaster risk intelligent identification and early warning method based on multi-element data fusion comprises a coal mine data access unit, a monitoring data visualization unit, an early warning information visualization unit, a model online incremental training unit, a model interface configuration unit, and an early warning rule configuration unit. The coal mine data access unit accesses the data of the coal mine monitoring sensor by configuring relevant parameters, queries the current and historical data of all accessed sensors in the coal mine in real time, supports filtering sensors of different locations and types, supports querying and displaying historical data according to different time ranges, and selects relevant monitoring indicators and their data for online incremental training of the model; The monitoring data visualization unit is used to display sensor information sources through an industrial computer display; The early warning information visualization unit is used to record the history of the coal mine major disaster risk intelligent identification and early warning method based on multi-element data fusion in the process of field application, and to filter and query historical early warning data by specifying time ranges and early warning levels; The model online incremental training unit is used to train and adjust the diffusion sample generation model, the MACN precursor risk identification model, and the MAKER method in the coal mine major disaster risk intelligent identification and early warning method based on multi-element data fusion, automatically records the last training data node, or manually sets a time range, selects a specified model for training, automatically saves the historical iteration version after training is completed, and transmits the results to the data platform, saves the training information to the database; The model interface configuration unit is used to configure model interface parameters, permission control, and model selection switching information, and to call interfaces such as diffusion model sample generation, precursor risk discrimination, and multi-element fusion early warning; The early warning rule configuration unit can manually configure risk early warning rules according to different situations on site, including the definition rules of no warning, first-level warning, and second-level warning in the early warning rules.

Claims

1. A method for intelligent identification and early warning of major disaster risks in coal mines based on multi-source data fusion, characterized in that, The steps are as follows: S1. Collect a large number of real samples of coal mine working face disasters to construct a time series dataset. Use a one-dimensional deep channel attention residual convolution-LSTM model based on self-supervised learning to identify abnormal data in the time series dataset. Use the multiple interpolation model MICE to interpolate missing data in the time series dataset to form a real sequence composed of data with no outliers in the indicators. S2. Perform Pearson correlation coefficient and Spearman rank correlation analysis on the preprocessed time series dataset to determine the monitoring and early warning indicators for major disaster risks in different coal mines. S3. Analyze the risk monitoring and early warning indicators of major coal mine disasters. Input the risk precursors, interference and noise signals of various coal mine working face hazards obtained from the analysis into the sample augmentation model DDPM based on denoising diffusion probability to expand the sample of different early warning indicators and their corresponding major coal mine disaster risks. S4. Using expanded early warning indicators and their corresponding samples of major coal mine disaster risks, a one-dimensional multi-scale multi-attention parallel temporal deep convolutional neural network (MACN) is trained to identify the precursor signals of major coal mine disaster risks and to distinguish and eliminate interference and noise signals. S5. Using the maximum likelihood method of evidence reasoning MAKER, construct a multi-indicator risk fusion identification model for various major disasters, and use the precursor signal identification results to complete the training of the multi-indicator risk precursor identification model for various major disasters. S6. Establish a major disaster risk intelligent identification and early warning system using a multi-indicator risk precursor identification model for major disasters, so as to realize intelligent application of coal mines and real-time monitoring and early warning of major disaster risks.

2. The intelligent identification and early warning method for major coal mine disaster risks based on multi-source data fusion according to claim 1, characterized in that, Real-world samples of coal mine working face hazards include data on multiple indicators: gas concentration, carbon monoxide concentration, coal dust concentration, fully mechanized mining resistance, surrounding rock delamination, anchor cable stress, temperature, wind speed, electromagnetic radiation, acoustic emission, long borehole and water flow.

3. The intelligent identification and early warning method for major coal mine disaster risks based on multi-source data fusion according to claim 2, characterized in that, The self-supervised learning-based one-dimensional deep channel attention residual convolution-LSTM model includes an encoder and a decoder. The output of the one-dimensional deep channel attention residual convolution in the encoder is connected to the input of the LSTM model, and the output of the LSTM model in the decoder is connected to the input of the one-dimensional deep channel attention residual convolution. The encoder takes a real sequence of data without outliers as input. It first generates an intermediate feature matrix, which is then input into the decoder. The decoder reconstructs the sequence based on the intermediate features. This process is repeated until a reconstructed sequence without outliers is obtained. Outliers can then be identified by passing a threshold. One-dimensional deep channel attention residual convolution is a fusion mechanism of temporal channel attention and frequency domain three-dimensional energy cross-gating based on one-dimensional residual convolution. This mechanism achieves highly sensitive modeling of sequence spectral changes. Combined with LSTM, it forms a self-supervised structure for multi-scale time-frequency joint modeling and accurate outlier identification of the temporal features of working surface indicators. Assuming a one-dimensional time series of indicators for monitoring and early warning of major disaster risks in coal mines is constructed, and using the real time series as input. ,in Where T is the batch size and T is the time step of the input sequence, the original sequence can be enlarged to the next dimension using the 1D convolution operation formula: C channels. ， in Let K be the weight of the c-th output channel relative to the input data at position j of the convolution kernel, and K be the kernel size. For bias units, output dimension C is the number of output channels. The convolution uses the same padding to keep the length of the output sequence unchanged. t is the time step position of the output after convolution. The convolutional backbone first uses two layers of convolution and non-linearity to extract feature information from the matrix Y output by the 1D convolution, which contains different receptive fields. ， in Let the kernel be a one-dimensional convolution kernel of length k and dilation rate d. , Using two one-dimensional convolutional kernels of different sizes, channel attention is calculated on the convolutional backbone branches, including channel compression and frequency domain cross-gating; first, the feature information is... Perform channel compression: ， ， ， in This represents the mean of the c-th channel over the time dimension. , It is a non-linear activation function. This represents the scaling factor for the c-th channel, generated by a fully connected + Sigmoid algorithm. , These represent the weight matrices of the fully connected layer, This is the feature vector of the c-th channel output by the main branch of the convolution; Perform a Fast Fourier Transform (rFFT) on the input data for each channel and take the amplitude: ， ， in Let f(x) be the FFT coefficient of the x-th channel at the x-th frequency. For Fourier transform, The amplitude is then used as the basis for determining the frequency index axis of each channel after the FFT, which is then divided into three index sets. Define frequency band energy: ， in The average energy vectors of the low, mid, and high frequency bands of the nth channel are obtained; then, the gating weights are obtained by sharing the average energy vectors of the low, mid, and high frequency bands using a multilayer perceptron. ， in , This is the weight matrix. For bias terms, The frequency domain gating factor generated for the nth channel; Finally, the gating weights are merged with the channel compression results: ， The feature vectors of the coal mine major disaster risk monitoring and early warning indicators after dimensional matching are added to the fusion result to form a residual connection: ， This is the feature vector after dimensional matching processing of the monitoring and early warning indicators for major coal mine disaster risks. The output of the first-layer depth channel attention residual convolution module is used as the input of the next layer; the number of layers needs to be set according to the complexity of different index fluctuations, and can be set between 3 and 8 layers. The encoder finally outputs a one-dimensional depth channel attention residual convolution with the dimension of... The data is input into the LSTM; the LSTM output is... Generate intermediate feature matrices; The input to the decoder's one-dimensional depth channel attention residual convolution is the output of the LSTM. ; The loss L1 is calculated based on the reconstructed sequence values ​​from intermediate features and the true sequence values ​​from the real, anomaly-free data. L1 is then used to evaluate the reconstructed sequence values ​​and reconstruction error generated by the LSTM model. ， in The reconstructed sequence values ​​output by the decoder. This is the actual sequence value input to the encoder.

4. The intelligent identification and early warning method for major coal mine disaster risks based on multi-source data fusion according to claim 3, characterized in that, The workflow of MICE based on the chain-like multi-interpolation model is as follows: Let the data variables in the time series data be represented as The MICE model first fills the missing values ​​of each variable with random values. In each subsequent iteration, it uses other variables in the time series to fill the missing values ​​of the current variable. Other variables include: gas concentration data, carbon monoxide concentration data, coal dust concentration data, fully mechanized mining resistance data, surrounding rock delamination data, anchor cable stress data, temperature data, wind speed data, electromagnetic radiation data, acoustic emission data, long borehole and water flow rate. For variables The The interpolated values ​​for the next iteration are obtained through the following distribution: ， in MICE model parameters at the next iteration The fully conditional posterior distribution, for The prior distribution; Calculate using posterior prediction distribution iteration Updated interpolation values: ， in For variables , Observed under the MICE model The probability, This represents the observed, non-missing sample values. This indicates the values ​​of all variables except the current variable. ,in It is an interpolated or updated value. It is an uninserted or unupdated value.

5. The intelligent identification and early warning method for major coal mine disaster risks based on multi-source data fusion according to claim 4, characterized in that, Analysis of pre-disaster data collected from major coal mine disasters using Pearson correlation coefficient and Spearman rank correlation coefficient revealed that gas disasters are correlated with gas concentration, coal dust concentration, fully mechanized mining resistance, surrounding rock delamination, anchor cable stress, temperature, wind speed, electromagnetic radiation, and acoustic emission. Fires are related to gas concentration, carbon monoxide concentration, temperature, and wind speed; dust is related to long boreholes and water flow in water hazards; and roof and other coal and rock dynamic disasters are correlated with gas concentration, coal dust concentration, fully mechanized mining resistance, surrounding rock delamination, anchor cable stress, electromagnetic radiation, and acoustic emission. Based on these findings, an index set for different major disasters was established.

6. The intelligent identification and early warning method for major coal mine disaster risks based on multi-source data fusion according to claim 5, characterized in that, The one-dimensional multi-scale, multi-attention parallel temporal deep convolutional neural network (MACN) captures features at different time scales using multi-scale convolutions. It generates branch attention, channel attention, and temporal attention based on lightweight weights to finely weight features. Finally, a residual network is used to reduce information loss from multiple layers. Downsampling and feature upscaling are performed on the input data to the MACN network: the input data is processed sequentially through a convolutional neural network, a batch normalization layer, and a ReLU activation function, and then the preprocessing result is output using a max pooling layer; The preprocessing results are input into a feature extraction module consisting of multiple layers of multi-scale convolutional attention blocks. The core process of each multi-scale convolutional attention block is as follows: the input data is mapped to a dimension using one-dimensional convolution, and then a branch is established for each convolutional kernel, with each branch outputting a set of branch vectors. ,in Output as a dimension mapping. For output channel size, The computational network for each branch includes one-dimensional convolution, batch normalization, and activation functions. Stack all branch data into a data tensor ,in Number of branches; Input the tensor into the branch attention and dynamically select the contribution of different receptive fields; Branch attention is first averaged over the time dimension, and then the branch attention weights are obtained through a multilayer perceptron (MLP) and softmax. Then, the weights of each branch are merged, as shown in the expression: ， ， in For branch attention results, For the slice vector of the b-th branch, the branch attention result is input into the channel attention function to amplify the channels that are informative for the task and suppress irrelevant channels. The channel attention function first performs global temporal pooling to obtain the feature descriptions of each channel, and then uses a linear layer-activation function-linear layer and softmax to obtain the channel attention weights. Finally, the attention weights for each channel are recalibrated, and the calculation formula is as follows: ， ， in The result of channel j being the Hadamard product of channel weights is then input into temporal attention. Temporal attention concatenates the features of all branches into a data tensor across all channels. Then, the temporal attention weights at each time point are implemented by combining one-dimensional convolution with the softmax function. Then, time-weighted processing is applied to each channel: ， ， Finally, the channel dimension is output through one-dimensional convolution and batch normalized projection, and then added to the residual. The calculation method is as follows: ， If it is not the last layer, it is used as the input for the next layer; otherwise, the classification result is output through global average pooling and a classification layer. ， The expanded datasets of each indicator are labeled with risk, interference, and noise signal types. These are then input into the MACN model for supervised training of a multi-classification recognition model for precursor signals. The trained MACN model effectively classifies dangerous, interference, and noise signals based on real-time monitoring of signal changes, filtering out dangerous signals and excluding other interference and noise signals.

7. The intelligent identification and early warning method for major coal mine disaster risks based on multi-source data fusion according to claim 1, characterized in that, A risk fusion identification model based on multiple indicators of various major disasters is used to identify the precursor risks of major coal mine disasters based on information from various sensors. The state space model and evidence space model are used to handle uncertainty through basic probability allocation, trust function and likelihood function. Dempster combination rule is used to fuse multi-source sensor data. The output of the multi-indicator risk fusion identification model for various major disasters represents the precursor risk probability information of each major disaster. Therefore, the discrimination results output by the multi-indicator risk fusion identification model for various major disasters are transformed into basic probability assignments, quantifying the risk status subset of each precursor. The level of support is: ， In the formula, This indicates that the i-th feature supports the basic probability distribution of a specific state. and The basic probabilities of risky and risk-free states within a subset of states; To measure the interdependence between different sensor information sources, a joint dependency index is introduced. : ， in It is a joint likelihood approximation, if This reflects a conflict; The joint dependency coefficient indicates that the higher the coefficient, the stronger the consistency between the two information sources on focal elements G and S. G and S represent subsets of states to be judged supported by evidence, including {risk}, {no risk}, or combinations thereof. Let be the joint likelihood measure of the information source i and the information source j. , Let G and S represent the likelihood of focusing elements G and S, respectively, for sensor information source i and sensor information source j. Let A, B, and E be the three sensor information sources. Given the basic probability distribution and interdependence coefficients... Under the given conditions, the joint support of the three sensor information sources is calculated using the following formula: ， in Let F be the joint dependency index of the three sensors, and F be the fused focal element, power set. for{ {Risk}, {No Risk}, {Risk, No Risk}}; the fusion contribution is only valid if the intersection of the three is not empty; for all combinations with empty intersections, their quality is considered conflict quality: ， Finally, the mass of all non-empty focal elements is normalized using the following formula: where the focal element is the aforementioned power set { A subset of {risk}, {no risk}, {risk, no risk}; ， Subset-level reliability weights for the three sensors The training process aims to minimize the error between the fusion result and the true label. ， in This is the s-th sample, including sensor data collected by sensors A, B, and E, and the fused basic probability assignment of the corresponding real disaster state labels. The label is the real label. The minimization problem is solved using the finite-memory quasi-Newton method L-BFGS-B. Finally, in the fused basic probability assignment, the single-element focal element with the highest support is selected as the predicted label. ， If the maximum support falls on the power set {risk, no risk}, it is considered a "fuzzy / uncertain" state, and the decision remains undetermined; This function outputs whether there is a risk or uncertainty regarding the current disaster.

8. The intelligent identification and early warning method for major coal mine disaster risks based on multi-source data fusion according to claim 7, characterized in that, Establish comprehensive disaster early warning rules to better achieve configurable and dynamic comprehensive early warning in complex coal mine conditions: 1) If no signal in the disaster indicator set is identified as indicating a precursor risk, it is considered that there is no risk of outbreak, and no risk will be marked, and no warning will be issued; 2) If, among the key indicators specified by experts, a single sensor information source is identified as having a precursor risk, then the current time will be marked as a potential precursor to a disaster risk, and a Level 1 warning will be issued; 3) If the identified precursory risk persists for more than 15 minutes, or if two or more sensor information sources with precursory characteristics of disaster risk are detected simultaneously among the key indicators specified by experts, or if the presence of risk is output according to the MAKER discrimination method, then the situation is marked as a level 2 warning.

9. A risk intelligent identification and early warning system based on the multi-source data fusion method for intelligent identification and early warning of major coal mine disaster risks as described in any one of claims 1-8, characterized in that, It includes a coal mine data access unit, a monitoring data visualization unit, an early warning information visualization unit, an online incremental training unit for models, a model interface configuration unit, and an early warning rule configuration unit; The coal mine data access unit can access data from coal mine monitoring sensors by configuring relevant parameters, query the current and historical data of all access sensors in the coal mine in real time, and support filtering sensors of different locations and types. It can also support querying and displaying historical data according to different time ranges, and select relevant monitoring indicators and their data to perform online incremental training of the model. The monitoring data visualization unit is used to display sensor information sources on an industrial computer monitor; The early warning information visualization unit is used to record the historical records of early warnings during the field application of the intelligent identification and early warning method for major coal mine disaster risks that integrates multiple data. The historical early warning data can be filtered and queried by specifying the time range and early warning level. The online incremental training unit is used to train and adjust the diffusion sample generation model, the MACN precursor risk identification model, and the MAKER method in the intelligent identification and early warning method for major coal mine disaster risks based on multi-data fusion. It automatically records the data nodes used in the last training, or allows manual setting of the time range, selection of a specific model, and training. After training is completed, it automatically saves the historical iterations and transmits the results to the data platform, saving the training information to the database. The model interface configuration unit is used to configure information such as model interface parameters, permission control and model selection and switching, and to call interfaces such as diffusion model sample generation, precursor risk identification and multi-element fusion early warning. The early warning rule configuration unit allows manual configuration of risk early warning rules based on different on-site conditions, including the definition rules for no early warning, Level 1 early warning, and Level 2 early warning.

Images