Dynamic threshold division method and multi-factor early warning method for gas emission of working face

Abstract

The present application belongs to the technical field of coal mine safety, and particularly relates to a dynamic threshold division and multi-factor early warning algorithm model for gas emission of a working face, a dynamic threshold division algorithm model, which comprises an FC-Kmeans model and a GA genetic algorithm model, the FC-Kmeans model comprises two stages: initialization of a clustering center and updating of the clustering center, the GA genetic algorithm model is a GA model based on a fitness function, and is used for clustering quality evaluation of a clustering result output by the FC-Kmeans model; a multi-factor algorithm model, which comprises a Transformer model and a MOA model, the Transformer model is composed of three parts of input, encoder structure and output, and embedding of the MOA model enables information entering the input part of the Transformer model to first enter a multi-head attention layer to extract information, and the multi-head attention layer is composed of multiple self-attention layers. The two models can dynamically divide a threshold range corresponding to a gas early warning level, and comprehensively early warn a real-time state, a fluctuation state and a trend state of gas emission of the working face.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of coal mine safety, and particularly relates to a dynamic threshold division and multi-factor early warning algorithm model for gas emission of a working face. BACKGROUND

[0002] Gas disaster is one of the main disasters faced by current underground coal mines. At the current stage, the monitoring and early warning of gas disaster in coal mines is mainly through the coal mine safety monitoring system, relying on AQ6201 and AQ1029 to arrange gas sensors at key points in the mine to monitor the gas concentration value in real time. The monitoring system generally monitors and warns the gas emission condition in real time according to the principle of overrun warning. On the basis of the time series of gas concentration, the existing technology establishes gas emission fluctuation anomaly analysis and gas emission trend anomaly analysis warning models and methods through the real-time concentration value and the moving average value of a certain period of time, respectively, to analyze and warn the gas emission from the real-time concentration value, fluctuation anomaly state and trend anomaly state of the gas emission. These methods improve the monitoring ability of gas emission anomalies by more deeply mining and analyzing the gas monitoring concentration value. However, the gas emission of the coal mine working face is affected by various factors, and the real-time concentration value change, fluctuation change and trend change of the gas emission are not isolated. How to establish a multi-factor gas early warning method to comprehensively warn the real-time state, fluctuation state and trend state of the gas emission of the working face is one of the important research contents of the current research on the early warning of the gas emission of the coal mine working face.

[0003] The existing clustering methods and their improved methods have wide applications in many fields, and they can reveal the internal structure of data and are more conducive to the extraction of subsequent feature signals. The existing clustering methods are affected by local optimal solutions, which reduces the stability and predictability of the algorithm. Secondly, the selection of the initial center of the KMeans clustering algorithm will affect the stability of the clustering results.

[0004] For the gas early warning problem of the working face, the research methods of machine learning include K-Nearest Neighbors (KNN), Support Vector Classifier (SVM), Naive Bayes, and Decision Tree classification.

[0005] The research method for deep learning considers that the multi-factor gas early warning is a special sequence-to-sequence type task, and some recent works begin to adopt the Transformer architecture to utilize its powerful sequence modeling capability. However, there are still some problems in the Transformer-based work. First, the attention mechanism always models the space-time dependence with the same parameters, even if the gas concentration will show different space-time patterns at different times. Second, in the gas early warning task of the Transformer model, the dimensions of the input and output information must match, which restricts the further improvement of the multi-factor gas early warning effect. SUMMARY

[0006] The present application aims to solve one of the technical problems existing in the prior art.

[0007] To this end, the present application provides a dynamic threshold division and multi-factor early warning algorithm model for gas emission of a working face.

[0008] The technical scheme adopted by the present application to solve its technical problems is:

[0009] The algorithm model for dynamic threshold division of gas emission of a working face comprises an FC-Kmeans model and a GA genetic algorithm model. The FC-Kmeans model comprises two stages: initialization of cluster centers and updating of cluster centers. The GA genetic algorithm model is a GA model based on a fitness function, which is used to evaluate the clustering quality of the clustering results output by the FC-Kmeans model.

[0010] By adopting the above technical scheme, an intelligent threshold division algorithm based on GF-KMeans is established, which dynamically divides the threshold range corresponding to the gas early warning level in combination with a genetic algorithm (GA) and an FC-KMeans algorithm. The FC-KMeans algorithm realizes more accurate control in the selection of initial cluster centers. The GA algorithm solves the local optimal problem in the traditional K-Means algorithm and enhances the stability and predictability of the algorithm.

[0011] Further, in the stage of initialization of cluster centers, a sample point is selected from the sample space as the first cluster center Secondly, for each sample point , the distance between it and the nearest cluster center is calculated.

[0012] Further, in the stage of updating of cluster centers, part of the cluster centers are set as fixed centers , and the rest are non-fixed centers The average distance between each candidate center and all fixed centers is calculated, and the maximum is selected. one As non-fixed initial clustering centers, the distance of each sample point from all clustering centers is calculated.

[0013] Further, the non-fixed clustering centers are updated, and the formula is as follows: wherein is the cluster determined by the clustering center .

[0014] Further, the fitness function uses the silhouette coefficient to evaluate the clustering quality of the clustering results output by the FC-Kmeans model.

[0015] Further, the fitness function performs multiple iterations on the output results of the FC-Kmeans model, and when there is no improvement in the last 10 consecutive iterations, the algorithm is terminated in advance.

[0016] A multi-factor algorithm model based on dynamic threshold division of working face gas emission, comprising a Transformer model and a MOA model, the Transformer model being composed of three parts of input, encoder structure and output, the embedding of the MOA model enabling information entering the input part of the Transformer model to first enter a multi-head attention layer to extract information, the multi-head attention layer being composed of multiple self-attention layers.

[0017] By adopting the above technical solution, on the basis of intelligent threshold division, a multi-factor gas warning model based on MOA-Transformer is established, the multi-head optimization attention mechanism is used to construct the feature representation of classification, and the encoder structure of the Transformer is used to perform gas concentration grade warning classification.

[0018] Further, the input data of the input part of the MOA model is processed by label embedding to become a group of index data in the basic index.

[0019] Further, the multi-head optimization attention mechanism is established by adding an intermediate matrix in the parameter learning matrix W, and three learnable parameter matrices W Q , W K , W V are used to multiply X i to obtain three representation vectors: , , , and the formula of the Transformer optimization attention mechanism is as follows:

[0020]

[0021]

[0022] wherein, α ij representing the calculation of the first output information, the weight of the first K j how much weight is allocated, representing the calculated weight score is reduced by a certain multiple, and the V j and α ij weighted sum of the value of the first A output information. i

[0023] Further, it is defined that wherein, h represents the number of heads, can be decomposed into h matrices, each corresponding to a .

[0024] The beneficial effects of the present application are, first of all, to establish an intelligent threshold division algorithm based on GF-KMeans, which combines genetic algorithm (GA) and FC-KMeans algorithm to dynamically divide the threshold range corresponding to the gas warning level. The GA algorithm solves the local optimal problem in the traditional K-Means algorithm, and enhances the stability and predictability of the algorithm. The FC-KMeans algorithm realizes more accurate control in the selection of initial clustering center. Then on the basis of intelligent threshold division, a multi-factor gas warning model based on MOA-Transformer is established, which constructs the feature representation of classification through multi-head optimization attention mechanism, uses the encoder structure of Transformer to perform gas concentration level warning classification, and can comprehensively warn the real-time state, fluctuation state and trend state of the gas emission of the working face. BRIEF DESCRIPTION OF DRAWINGS

[0025] The present application will be further described below in conjunction with the drawings and examples.

[0026] Figure 1 is a structural schematic diagram of the GF-KMeans model in the present application.

[0027] Figure 2 is a structural schematic diagram of the MOA-Transformer model in the present application.

[0028] Figure 3is a gas concentration real-time data curve chart in the embodiment of the present application.

[0029] Figure 4 is a gas concentration 5min moving average data curve chart in the embodiment of the present application.

[0030] Figure 5 is a gas concentration 5min maximum value data curve chart in the embodiment of the present application.

[0031] Figure 6 is a gas concentration 10min moving average data curve chart in the embodiment of the present application.

[0032] Figure 7 is a gas concentration 10min maximum value data curve chart in the embodiment of the present application.

[0033] Figure 8 is a gas concentration 30min moving average data curve chart in the embodiment of the present application.

[0034] Figure 9 is a use flowchart of the threshold value division algorithm model and the early warning algorithm in the present application. DETAILED DESCRIPTION

[0035] The application will now be described in further detail with reference to the drawings. These drawings show only the basic structure of the application and therefore only show the components relevant to the application, in a simplified manner. The application can therefore vary in its implementation.

[0036] In the description of the present application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only used to facilitate the description of the present application and simplify the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise stated, the meaning of "a plurality of" is two or more.

[0037] In the description of the present application, it is necessary to point out that, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "linking" should be understood in a broad sense, for example, it can be fixed connection, or detachable connection, or integrally connected; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium, or the internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0038] An algorithm model for dynamic threshold division of gas emission of working face, an intelligent threshold division algorithm based on GF-KMeans is proposed in this paper, referring to Figure 1 , the algorithm combines genetic algorithm (GA) and FC-KMeans algorithm. First, the GA algorithm solves the local optimal problem in the traditional K-Means algorithm, and enhances the stability and predictability of the algorithm. Secondly, the FC-KMeans algorithm realizes more accurate control in the selection of initial cluster centers.

[0039] FC-KMeans algorithm is an improved clustering algorithm, which introduces the concept of fixed center on the basis of standard KMeans algorithm, so that part of the cluster centers can be fixed and not updated. This algorithm can more flexibly control the change of cluster centers in the clustering process. The algorithm mainly consists of two stages: initialization of cluster centers and updating of cluster centers.

[0040] In the stage of initializing cluster centers, first, a sample point is randomly selected from the sample space as the first cluster center . Secondly, for each sample point , calculate its distance from the nearest cluster center , the formula is:

[0041] (1)

[0042] Then, select the sample point as the next cluster center with probability , where . By repeating these steps, until initial cluster centers are selected. In this way, the dispersion of the initial cluster centers can be ensured, which helps to avoid local optimal solution.

[0043] In the stage of updating cluster centers, part of the cluster centers are set as fixed centers , and the rest are non-fixed centers . For each candidate center , calculate the average distance from all fixed centers , the formula is as follows:

[0044] (2)

[0045] Select the maximum One As the initial non-fixed clustering center. For each sample point , calculate its distance from all clustering centers, and assign it to the nearest cluster:

[0046] (3)

[0047] Finally, update the non-fixed clustering center accordingly, the formula is as follows:

[0048] (4)

[0049] Where is the cluster determined by the clustering center .

[0050] The GF-KMeans algorithm is a fusion algorithm of genetic algorithm (GA) and FC-KMeans clustering algorithm. Genetic algorithm (GA) is an optimization method that simulates the process of natural selection, which can find the best solution in the entire search space through selection, crossover, mutation and replacement operations. After the FC-KMeans model clustering is completed, the GA model is run to evaluate the iteration quality until the set requirements are met.

[0051] First, initialize the parameters of genetic algorithm, including the maximum number of iterations, population size, etc. Second, create a GA model based on the fitness function, where the fitness function uses the silhouette coefficient to evaluate the clustering quality.

[0052] Silhouette Score (Silhouette Coefficient): This index measures the tightness within the cluster and the separation between clusters. The value of the silhouette coefficient ranges from [-1, 1], and the higher the value, the better the clustering effect.

[0053] (5)

[0054] Where, is the average distance of sample from other samples in the same cluster, is the average distance of sample from the nearest cluster sample.

[0055] In the initialization phase, optimization is needed by running the GA model, and the fitness is evaluated to search for the best clustering center. In this process, the output results of the FC-KMeans model need to be iterated multiple times. When there is no improvement in continuous 10 iterations, the algorithm is terminated in advance. The silhouette coefficient S The calculation formula of the evaluation is as follows: n is the number of iterations:

[0056] (6)

[0057] A multi-factor early warning algorithm model based on dynamic threshold division of working face gas emission, the application proposes a multi-factor gas early warning model based on MOA-Transformer, first uses the encoder structure of Transformer to perform gas concentration level early warning classification. Secondly, the multi-head optimization attention mechanism is used to construct the feature representation of classification.

[0058] Reference Figure 2 , the Transformer model is composed of input, encoder structure and output. As shown in the following figure, the red circle is the encoder structure.

[0059] The input part of the Transformer is the result of adding Input Embedding and Positional Embedding.

[0060] The encoder structure of the Transformer is composed of multiple layers. The information input is first entered into the multi-head attention layer, which is composed of multiple self-attention layers. The extracted feature information enters the Add&Norm layer, where Add represents residual connection (Residual Connection) to prevent grid degradation, and Norm represents Layer Normalization to normalize the activation function of each layer. Feed Forward represents the feed-forward fully connected layer, which maps the data to a high latitude space and then to a low latitude space through linear transformation, extracting deeper features. Finally, it is processed through the Add&Norm layer and ends.

[0061] The output part of the Transformer is to output the feature information processed by the encoder to the Linear layer and the Softmax layer, and obtain the final classification information after processing.

[0062] The self-attention mechanism of the Transformer is a mechanism that focuses on local features. The MOA model processes the input part, the basic index data, into , to represent a group of index data in the basic index.

[0063] In the self-attention model, the weight score is calculated by the dot product of Q and K, which can find the relationship between different parts of the information to construct the representation of the information. But in the gas warning task, the dimension of the input information does not match the dimension of the output information, which leads to the loss of important local feature information. Therefore, an optimized attention mechanism is proposed: a learnable parameter matrix as an intermediate matrix.

[0064] Three learnable parameter matrices W Q , W K , W V are multiplied by X i to obtain three representation vectors: , , . The formula of the optimized attention mechanism of the Transformer is as follows:

[0065] (7)

[0066] (8)

[0067] where, α ij represents how much weight is allocated to the th input information when calculating the th output information. K j represents that the calculated weight score is reduced by a certain multiple to prevent the gradient produced by Softmax from being too small, d k represents the length of the vector, and the role of the Softmax activation function is to adjust the calculated weight score to a reasonable range. The V j and α ij weighted sums are obtained. A i value of the th output information.

[0068] The multiple calculations using different parameters are added to obtain the multi-head optimized attention mechanism, which is defined as:

[0069] (9)

[0070] where h represents the number of heads, ​Decomposed into h matrices, each corresponding to one .

[0071] Embodiment 1

[0072] Taking the gas monitoring of a fully mechanized coal mining face in a high-gas mine in southwest China as an example, the mine adopts a work mode of two production shifts and one maintenance shift. This paper takes the monitoring data of the gas sensor on the working face as the basic data set for research. The mine has installed a KJ95X safety monitoring system to monitor the safety-related parameters of the mine in real time. The gas sensor on the working face has a patrol inspection time of about 20s, and about 2160 data sets of gas concentration values are selected for 12h a day. For the six indicators of the gas concentration on the working face: real-time value, five-minute moving average value, five-minute maximum value, ten-minute moving average value, ten-minute maximum value, and thirty-minute moving average value, the GF-KMeans algorithm is used to dynamically divide the threshold range corresponding to the gas warning level. The data is divided into five gas warning levels, from low to high, corresponding to four blue warnings, three yellow warnings, two orange warnings, one red warning, and normal state. Figures 3-8 The first 2160 data sets of this experiment are shown, reflecting the degree of change in gas concentration at different times.

[0073] Reference Figure 9 , based on index data, the real-time data and historical data monitored are input into the initialized model to train and update the model. Finally, the GF-KMeans-based intelligent threshold division algorithm model and the MOA-Transformer-based multi-factor gas warning model are obtained, and the dynamic threshold division results and multi-factor warning level division results are output.

[0074] 1.1 GF-KMeans-based intelligent threshold division algorithm experiment

[0075] 1.1.1 Intelligent threshold division algorithm performance experiment

[0076] In the performance experiment of the GF-KMeans intelligent threshold division algorithm, this paper selects three important evaluation indicators: Silhouette Score (Silhouette Score), Calinski Harabasz Score (CH Index), and Davies Bouldin Score (DB Index) to measure the clustering effect.

[0077] Calinski Harabasz Score (CH Index): This index evaluates the ratio of the dispersion between clusters to the dispersion within clusters. The larger the value, the better the clustering effect.

[0078] (10)

[0079] where is the number of training samples, is the number of classes, is the covariance matrix between classes, is the data covariance matrix within classes, is the trace of a matrix.

[0080] Davies Bouldin Score (DB Index): This index measures the ratio of the tightness within clusters to the separation between clusters, and the lower the value, the better the clustering effect.

[0081] (11)

[0082] where, is the number of clusters, is the average distance of all points in the th cluster to the cluster center. is the distance between the th and the th cluster center.

[0083] Comparison of three different intelligent threshold division algorithms: traditional KMeans, fixed center KMeans (FC-KMeans), and GF-KMeans combined with genetic algorithm. The GF-KMeans algorithm in this paper uses genetic algorithm to further optimize the fixed center. The key parameters of the experiment include population size of 30, cluster number of 5, and iteration number of 20. The experimental results are shown in the following table:

[0084] Table 1 Comparison of clustering algorithm performance

[0085]

[0086] From Table 1, we can clearly see that GF-KMeans is better than FC-KMeans and traditional KMeans method in all three evaluation indicators. The silhouette coefficient of GF-KMeans is 0.7754, which is significantly higher than the other two methods, indicating that GF-KMeans performs well in terms of tightness within clusters and separation between clusters. In terms of CH index, GF-KMeans also achieved the highest score of 4.8K, reflecting that GF-KMeans achieved a better balance between the dispersion between clusters and the dispersion within clusters. In terms of DB index, the score of GF-KMeans is 0.4152, which is lower than other methods, indicating that its clustering is more clear and clear. These outstanding advantages prove the effectiveness and feasibility of GF-KMeans in clustering analysis.

[0087] These results not only demonstrate the advantages of GF-KMeans in statistical indicators, but also reveal its potential in practical applications. Compared with traditional KMeans, GF-KMeans combines the global search ability of genetic algorithm, avoiding falling into local optimum, and finding more representative clustering centers. In addition, compared with FC-KMeans, GF-KMeans realizes more flexible and accurate clustering by dynamically adjusting the center position.

[0088] 1.1.2 Multi-factor threshold division experiment

[0089] The experiments in this paper are based on the six indicators in Table 2 to analyze the warning levels of the data, and use the GF-KMeans algorithm to determine the corresponding warning threshold range. The method in this paper can accurately capture these ranges and generate the following experimental results:

[0090] Table 2 Multi-factor threshold division results

[0091]

[0092] 1.2 MOA-Transformer algorithm experiment

[0093] This paper uses the intelligent threshold division results mentioned above as training data, including six indicators, namely real-time value, five-minute moving average, five-minute maximum, ten-minute moving average, ten-minute maximum, and thirty-minute moving average, to establish a multi-factor gas warning model to jointly determine the warning level of the working face.

[0094] The experiment calculates the warning level of each indicator according to the warning range, and calculates the total warning level by combining the weight and warning level. The MOA-Transformer algorithm experiment results are shown in Table 3.

[0095] Table 3 MOA-Transformer algorithm experiment results

[0096]

[0097] Table 3 shows the results of the gas concentration warning level. From the data in the table, it can be seen that the gas concentration data has complexity and variability, fully reflecting the complexity and variability of the data in the complex coal mine underground. The accuracy and robustness of the method in capturing complex and variable data.

[0098] This paper uses accuracy (Accuracy), precision (Precision), recall (Recall), and Macro-F1 as performance indicators:

[0099] Accuracy: The percentage of correctly predicted samples in the total, the specific formula is as follows:

[0100] (14)

[0101] where, is a true positive class, i.e., an instance is positive and is predicted as positive; is a false negative class, i.e., an instance is positive but is predicted as negative; is a false positive class, i.e., an instance is negative but is predicted as positive; is a true negative class, i.e., an instance is negative and is predicted as negative.

[0102] Precision: is an evaluation index for the predicted results. In the results predicted by the model as positive samples, the percentage of true positive samples.

[0103] (15)

[0104] Recall: is an evaluation index for the original samples. In the actual positive samples, the percentage of predicted positive samples.

[0105] (16)

[0106] Macro-F1: is a kind of harmonic mean, suitable for multi-classification tasks.

[0107] (17)

[0108] In this paper, four representative algorithms are selected for comparison experiments, including: recurrent neural network (RNN), convolutional neural network (CNN), long short-term memory network (LSTM), Transformer network and MOA-Transformer. The experimental results are shown in Table 4.

[0109] Table 4 Comparison of classification algorithm performance

[0110]

[0111] Table 4 is the comparison of accuracy, precision, recall, Macro-F1 of recurrent neural network, convolutional neural network, long short-term memory network, Transformer network and MOA-Transformer. It can be clearly seen from this paper that MOA-Transformer is better than recurrent neural network, convolutional neural network, long short-term memory network and Transformer network in the four evaluation indexes. The accuracy of the multi-factor gas warning model of MOA-Transformer is 86.17%, which is 3.45% higher than that of the Transformer algorithm; the precision of MOA-Transformer is 88.78%, which is 3.75% higher than that of the Transformer algorithm; the recall of MOA-Transformer is 85.23%, which is 4.70% higher than that of the Transformer algorithm; the Macro-F1 index of MOA-Transformer is 86.96%, which is 4.39% higher than that of the Transformer algorithm, which fully illustrates the superiority of MOA-Transformer algorithm in the gas concentration warning task.

[0112] In summary, first, an intelligent threshold division algorithm based on GF-KMeans is established, which dynamically divides the threshold range corresponding to the gas warning level by combining genetic algorithm (GA) and FC-KMeans algorithm. The GA algorithm solves the local optimal problem in the traditional K-Means algorithm, and enhances the stability and predictability of the algorithm. The FC-KMeans algorithm realizes more accurate control of the selection of initial cluster centers.

[0113] Then on the basis of intelligent threshold division, a multi-factor gas warning model based on MOA-Transformer is established, which constructs the feature representation of classification through multi-head optimization attention mechanism, and uses the encoder structure of Transformer to perform gas concentration level warning classification. Six basic indexes of gas concentration real-time value, five-minute concentration moving average value, five-minute concentration maximum value, ten-minute concentration moving average value, ten-minute concentration maximum value, and thirty-minute concentration moving average value are established. On this basis, the intelligent threshold division and multi-factor gas warning research work are carried out, which can comprehensively warn the real-time state, fluctuation state and trend state of the gas emission of the working face.

[0114] Based on the above ideal embodiments according to the present application, through the above description, relevant personnel can make various changes and modifications without deviating from the technical idea of the present application. The technical scope of the present application is not limited to the contents in the specification, and must be determined by the scope of the claims.

Claims

1. A method for dynamically classifying the threshold of gas outburst at a working face, characterized in that, It includes an FC-Kmeans model and a GA genetic algorithm model. The FC-Kmeans model includes two stages: initializing cluster centers and updating cluster centers. The GA genetic algorithm model is a GA model based on a fitness function, which is used to evaluate the clustering quality of the clustering results output by the FC-Kmeans model. The above algorithm is used to dynamically divide the threshold range corresponding to the gas warning level. In the initialization phase of cluster centers, a sample point is selected from the sample space as the first cluster center. Secondly, for each sample point Calculate its relationship with the nearest cluster center. The distance; During the stage of updating cluster centers, some cluster centers are set as fixed centers. The remainder are non-fixed centers. Calculate each candidate center With all fixed centers Choose the largest average distance. indivual As non-fixed initial cluster centers, each sample point is calculated. , its distance from all cluster centers; The formula for updating non-fixed cluster centers is as follows: ,in It is the cluster center Determined clustering.

2. The dynamic threshold division method for gas outburst at the working face according to claim 1, characterized in that, The fitness function uses the silhouette coefficient to evaluate the clustering quality of the clustering results output by the FC-Kmeans model.

3. The dynamic threshold division method for gas outburst at the working face according to claim 2, characterized in that, The fitness function iterates the output of the FC-KMeans model multiple times. If there is no improvement after 10 consecutive iterations, the algorithm terminates prematurely.

4. A multi-factor early warning method based on the dynamic threshold division method for gas outburst at the working face as described in any one of claims 1-3, characterized in that, It includes the Transformer model and the MOA model. The Transformer model consists of three parts: input, encoder structure, and output. The embedding of the MOA model allows the information entering the input part of the Transformer model to first enter the multi-head attention layer to extract information. The multi-head attention layer consists of multiple self-attention layers.

5. The multi-factor early warning method according to claim 4, characterized in that, The MOA model transforms the input data into a set of indicator data representing the basic indicators through label embedding processing. .

6. The multi-factor early warning method according to claim 5, characterized in that, A multi-head optimization attention mechanism is established by incorporating a parameter learning matrix W and an intermediate matrix, using three learnable parameter matrices. W Q , W K , W V Matrix multiplication X i We obtain three representation vectors: , , The formula for optimizing the attention mechanism in Transformer is as follows: in, α ij Represents the calculation of the first When outputting the first piece of information, give the first... Input information K j How much weight to assign? This indicates that the calculated weighted score has been reduced by a certain factor. d k express The length of the vector express The transpose of a vector, corresponding to all inputs V j and α ij Weighted summation yields the first Output information A i value.

7. The multi-factor early warning method according to claim 6, characterized in that, Defined as Where h represents the number of heads. It can be decomposed into h matrices, each matrix corresponding to a... .

Images