Intelligent coal mine gas outburst prediction method fusing priori knowledge segmented training

Abstract

The invention discloses a coal mine gas outburst intelligent prediction method fusing prior knowledge segmented training, and belongs to the technical field of gas early warning data processing. The method comprises the following steps: firstly, normalizing gas outburst prediction parameter data to form a standard sample; then, gas outburst sensitive indexes serve as priori knowledge, a sensitive feature extractor is constructed, sensitive index features are obtained from samples, model dimensions are kept unchanged, an MLP network is constructed for first-stage training, weight coefficients of the sensitive features and outburst risks are independently trained, and the trained weight coefficients are stored; and secondly, a new MLP network model is built for two-stage training, the stored weight coefficient is used as an initial weight to be assigned to the model, all indexes are trained to obtain a sensitive index weighted gas outburst prediction model, and coal mine gas outburst prediction is realized by using the trained model. According to the method, higher weight coefficients are given to sensitive indexes by fusing prior knowledge, and accurate prediction of coal mine gas outburst is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of gas early warning data processing, and in particular to a coal mine gas outburst intelligent prediction method integrating prior knowledge segmentation training. Background Art

[0002] With the improvement of the mechanization level of coal mining enterprises and the continuous advancement of smart mines, the safety level of coal mines has been greatly improved. However, with the increase in coal mining depth and the continuous extension of mining levels, the mining environment faced by coal mine production has become more complex, and various disasters still occur from time to time. Among various coal mine disasters, coal and gas outbursts remain one of the main coal mine disasters due to factors such as the complex and difficult to predict disaster-causing mechanisms and the serious consequences of accidents. Coal mine gas outburst prediction has always been a hot topic of research among experts and scholars. Various intelligent data processing methods have been applied to gas outburst prediction. However, the accuracy of the prediction model is highly dependent on the quality of the original data, and the selection of the initial value of the model also has a significant impact on the prediction results. Only processing the original data and improving the model prediction algorithm cannot obtain satisfactory gas outburst prediction results.

[0003] Coal mines accumulate a wealth of prior knowledge during production. Depending on the conditions in different mining areas, indicators such as K1 gas analysis and drill cuttings volume are highly correlated with gas outburst prediction results and are therefore sensitive indicators. Leveraging this prior knowledge and effectively incorporating it into prediction models is crucial for predicting gas outbursts. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent prediction method for coal mine gas outburst by integrating prior knowledge segmented training, which gives sensitive indicators a higher weight coefficient by integrating prior knowledge, thereby achieving accurate prediction of coal mine gas outburst.

[0005] To achieve the above objectives, the present invention provides a coal mine gas outburst intelligent prediction method integrating prior knowledge segmented training, comprising the following steps:

[0006] Step S1, collecting gas outburst prediction related indicator data;

[0007] Step S2: performing dimensionless normalization processing on the acquired data using the initial value method to form sample data for gas outburst prediction;

[0008] Step S3: construct a feature extractor based on prior knowledge, and use the feature extractor to process the predicted sample data to obtain a sensitive indicator feature data set;

[0009] Step S4: construct a one-stage MLP network and perform model training on the sensitive indicator feature data set to extract the model network weight coefficient;

[0010] Step S5: construct a two-stage MLP network and assign the network weight coefficient obtained in step S4 as the initial weight to the two-stage MLP network model;

[0011] Step S6: training the two-stage MLP network model based on the normalized prediction samples, and generating a prediction model for coal mine gas outburst prediction after training;

[0012] Step S7: Use the prediction model obtained in step S6 to predict coal mine gas outbursts, input the collected gas index parameter data into the prediction model, and the prediction model automatically gives the corresponding prediction results under the set danger level.

[0013] Preferably, in step S1, the gas outburst prediction-related indicator data include 9 types of training data, including coal seam gas content, K1 gas analysis amount, drill cuttings amount, excavation speed, distance from geological structure zone, coal gas release initial velocity, coal seam burial depth, coal seam thickness, and coal damage type, as well as the gas outburst hazard level corresponding to each set of data;

[0014] The hazard levels are divided into three levels: safe, general, and major, represented by 0, 0.6, and 1 respectively.

[0015] Preferably, in step S2, the initial value method is calculated as follows:

[0016]

[0017] Among them, p is the evaluation object; q is the evaluation parameter; x′ p (q) is the evaluation index after initialization; x p (q) is the evaluation index before initialization; x1(q) is the first data of the evaluation index.

[0018] Preferably, in step S3, the feature extractor operates as follows:

[0019] H sensitive =H original *T;

[0020]

[0021]

[0022] Among them, H sensitive is the sensitive index feature matrix obtained after processing by the feature extractor, H original is a matrix with n rows and m columns, where n is the number of sample data, m is the dimension of sample data, and H original Each prediction index is defined as X1 to X9, X nm H originalThe elements in , T is the sensitive indicator feature extractor, and two sensitive indicators are selected as prior knowledge and placed in the second and third columns of the sensitive indicator feature extractor T.

[0023] Preferably, the loss function of the segmented training MLP network model is as follows:

[0024]

[0025] Among them, Y i is the label of the first stage training sample; is the prediction result of the first stage training sample; Y j is the label of the second stage training sample; Prediction results for the second stage training samples; LOSS 全局 is the global loss; L 阶段1 is the loss in the first stage; L 阶段2 is the loss of the second stage; i and j are the indexes of the samples.

[0026] Therefore, the present invention adopts the above-mentioned intelligent prediction method for coal mine gas outburst that integrates prior knowledge and segmented training, and its technical effects are as follows: by integrating prior knowledge to give sensitive indicators a higher weight coefficient, accurate prediction of coal mine gas outburst is achieved.

[0027] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of an intelligent prediction method for coal mine gas outbursts that integrates prior knowledge and segmented training according to the present invention;

[0029] Figure 2 is the target neural network structure diagram;

[0030] Figure 3 Comparison of model performance under different training methods; Figure 3 (a) shows the changes of acc and loss without introducing prior knowledge; Figure 3 (b) The changes in acc and loss of one stage after integrating prior knowledge; Figure 3 (c) shows the changes in acc and loss in the second stage after integrating prior knowledge;

[0031] Figure 4 It is the prediction result of the neural network model without introducing prior knowledge on the test set;

[0032] Figure 5 The prediction results of the neural network model that integrates prior knowledge on the test set;

[0033] Figure 6 The confusion matrix of the model without introducing prior knowledge;

[0034] Figure 7 Confusion matrix of the prior knowledge fusion model. DETAILED DESCRIPTION

[0035] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0036] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0037] Example 1

[0038] like Figure 1 FIG. 1 is a flow chart of an intelligent prediction method for coal mine gas outbursts integrating prior knowledge segmented training according to the present invention, which specifically includes the following steps:

[0039] (1) Construct the characteristic matrix H of the drilling cuttings gas desorption index K1 and the drilling cuttings amount S sensitive . Create a sensitive indicator feature extractor T, H sensitive It can be obtained by the following formula:

[0040] H sensitive =H original *T (1);

[0041] H original The sample data obtained through grey correlation analysis include X1 (coal seam burial depth), X4 (drill cuttings gas desorption index K1), X5 (drill cuttings volume), X7 (gas concentration), X8 (gas content), and X9 (geological structure distance). original It is a matrix with 95 rows and 6 columns, where the sensitive indicator drill cuttings gas desorption index K1 is in the second column and the drill cuttings amount S is in the third column.

[0042]

[0043] (2) According to LOSS 全局 Function guide, with H sensitive For training samples, the model is trained and the index weight coefficient matrix and bias matrix (ω sensitive , b sensitive ).according to Figure 2 As shown, from the hidden layer to the output layer Taking the weight coefficient as an example, the weight update process is explained.

[0044] ①Forward process.

[0045] Similarly, we can find neth 12 、neth 13 to neth 1n .

[0046] outh 11 =f1(neth 11 ), where f1 is the first layer activation function. Similarly, outh 12 、outh 13 to outh 1n .

[0047] Similarly, we can find neth 22 、neth 23 to neth 2m .

[0048] outh 21 =f2(neth 21 ), where f2 is the second layer activation function. Similarly, outh 22 、outh 23 to outh 2m .

[0049] Similarly, you can ask for net O2 、net O3 、net O4 、net O5 .

[0050] out O1 =f3(net O1 ), where f3 is the activation function of the third layer. Similarly, out O2 、out O3 、out O4 、out O5 . out O1 to out O5 The measured value

[0051] in, is the weight coefficient of the lth layer, s represents the index of the neuron, and w represents the index of the neuron in the previous layer; b s is the bias term of the sth neuron; neth ls is the net input of the sth neuron in the lth layer; outh ls is the output of the sth neuron in the lth layer, obtained by applying the activation function to the net input; net Os is the net input of the sth neuron in the output layer; net Os is the output of the sth neuron in the output layer; f lis the activation function of the lth layer.

[0052] ②Reverse process.

[0053] by Take weight update as an example:

[0054]

[0055] Where η is the learning rate, k is the number of iterations, and Δ represents the amount of change. In this model framework, during the first phase of training, the second phase of training has not yet begun, so:

[0056]

[0057] Applying chain derivation, formula (4) is transformed into:

[0058]

[0059] make real O1 is the true value of the sample label. Similarly, we can get E O2 to E O5 .

[0060] L 阶段1 =E O1 +E O2 +E O3 +E O4 +E O5 (6);

[0061] Solve the first factor of equation (6):

[0062]

[0063] The result of formula (8) can be obtained by direct settlement through the forward process.

[0064] Solve the second factor of equation (6):

[0065]

[0066] According to the f3 function type, the derivative can be directly obtained, and then the forward process can be used to directly calculate the result.

[0067] Solve the third factor of equation (6):

[0068]

[0069] The results can be obtained directly using the forward process.

[0070] Formula (6) can be further transformed into:

[0071]

[0072] So far, all three factors in formula (6) are directly calculated by the forward process.

[0073] Substituting formula (6) into formula (4), we get Weight update formula:

[0074]

[0075] Similarly, other weight coefficient iteration methods can be obtained, which will not be described here.

[0076] (3) sensitive , b sensitive ) are the initial weight coefficients and bias parameters, with H original The second stage of model training is performed for the training samples, using gradient descent to obtain the final prediction model. The weight update process in this stage is similar to that in the first stage. When the partial derivative of the loss function with respect to the weight ω is taken, the loss value in the first stage is constant. The weight update process is not detailed here.

[0077] Experimental results.

[0078] The neural network model uses a three-layer structure, with five neurons per layer. The epoch number is set to 1000. Accuracy, Precision, Recall, and f1_score (Equations 12 to 15) are used as evaluation metrics. Given that the coal and gas outburst prediction model has five prediction levels and is a multi-classification problem, Accuracy, Precision, Recall, and f1_score require a comprehensive consideration of the overall performance of each category.

[0079]

[0080] Where N correct Indicates the number of samples classified correctly in all categories, N total Indicates the total number of samples.

[0081]

[0082] Where M is the number of sample categories. In this embodiment, M is the number of levels of coal and gas outburst prediction in the test sample. v Indicates the proportion of the accuracy of a certain category in the overall accuracy, α v It is an adjustable parameter, which depends on the specific situation. In this embodiment, it is considered that the importance of coal and gas outburst prediction level is equal, so α can be set to v =1 / M. TP v Indicates the number of correct predictions among samples of a certain category predicted as positive examples, FP vIndicates the number of incorrect predictions in the sorted samples for this category.

[0083]

[0084] Where, FN v The number of samples predicted incorrectly as negative examples for a certain category.

[0085]

[0086] Where, P v Indicates the accuracy of a certain category, R v represents the recall rate of the category, and v represents the index variable.

[0087] The loss function loss and accuracy acc of the neural network model without introducing prior knowledge are compared with the loss function loss and accuracy acc of the neural network model integrating prior knowledge, such as Figure 3 .

[0088] Figure 3 (a) and (b) in the figure are both initial training of the data. It can be seen that Figure 3 The acc curve of (b) has better convergence effect, which is because Figure 3 The training data used in (b) is the sensitive indicator data extracted by the sensitive indicator feature extractor. Figure 3 (a) The training data used is less noisy. Figure 3 (c) shows the loss function and accuracy curve of the two-stage training model integrating prior knowledge. This model obtains the weight parameters of sensitive indicator training, namely Figure 3 The training of (c) is carried out under the guidance of prior knowledge. It can be seen that only 40 steps of training are needed to reach convergence.

[0089] The prediction results of the neural network model without prior knowledge on the test set are as follows: Figure 4 , the prediction results of the neural network model integrating prior knowledge on the test set are as follows Figure 5 .

[0090] The Accuracy, Precision, Recall, and f1_score results of the model that integrates prior knowledge and the model that does not introduce prior knowledge are shown in Table 1.

[0091] Table 1 Comparison of Accuracy, Precision, Recall, and f1_score between models integrating prior knowledge and models without prior knowledge

[0092]

[0093] from Figure 4 and Figure 5 The comparison of the model evaluation indicators in Table 1 shows that, under the same neural network model parameter conditions, the neural network model integrating prior knowledge outperforms the neural network model without prior knowledge in terms of accuracy, precision, recall, and f1_score. It can be concluded that the introduction of prior knowledge has improved the performance of the coal and gas outburst prediction model.

[0094] By comparing the confusion matrix Figure 6 and 7 , combined with Figure 4 、 5 The prediction results show that there are no samples with a prominence risk level of 3 in the real test samples. The model without prior knowledge incorrectly identified a prominence risk level of 3, resulting in a total of five predictions. The model incorporating prior knowledge did not misjudge the prominence risk level of 3, resulting in a total of four predictions, indicating that the model incorporating prior knowledge is more consistent with the actual situation. Furthermore, the confusion matrix shows that the model incorporating prior knowledge performs better for high prominence risk levels (risk levels 2, 3, and 4), where there are fewer sample labels.

[0095] Therefore, the present invention adopts the above-mentioned coal mine gas outburst intelligent prediction method of segmented training integrating prior knowledge, and gives sensitive indicators a higher weight coefficient by integrating prior knowledge, thereby achieving accurate prediction of coal mine gas outburst.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent prediction method for coal mine gas outburst integrating prior knowledge segmented training, characterized in that: The following steps are involved: Step S1, collecting gas outburst prediction related indicator data; Step S2: performing dimensionless normalization processing on the acquired data using the initial value method to form sample data for gas outburst prediction; Step S3: construct a feature extractor based on prior knowledge, and use the feature extractor to process the predicted sample data to obtain a sensitive indicator feature data set; Step S4: construct a one-stage MLP network and perform model training on the sensitive indicator feature data set to extract the model network weight coefficient; Step S5: construct a two-stage MLP network and assign the network weight coefficient obtained in step S4 as the initial weight to the two-stage MLP network model; Step S6: training the two-stage MLP network model based on the normalized prediction samples, and generating a prediction model for coal mine gas outburst prediction after training; Step S7: Use the prediction model obtained in step S6 to predict coal mine gas outbursts, input the collected gas index parameter data into the prediction model, and the prediction model automatically gives the corresponding prediction results under the set danger level.

2. The method for intelligent prediction of coal mine gas outbursts integrating prior knowledge and segmented training according to claim 1 is characterized in that: In step S1, the gas outburst prediction-related indicator data includes nine types of training data: coal seam gas content, K1 gas analysis volume, drill cuttings volume, tunneling speed, distance from geological structure zone, coal gas release initial velocity, coal seam burial depth, coal seam thickness, and coal damage type, as well as the gas outburst hazard level corresponding to each set of data; The hazard levels are divided into three levels: safe, general, and major, represented by 0, 0.6, and 1 respectively.

3. The method for intelligent prediction of coal mine gas outburst integrating prior knowledge segmented training according to claim 1 is characterized in that: In step S2, the initial value method is calculated as follows: Among them, p is the evaluation object; q is the evaluation parameter; x′ p (q) is the evaluation index after initialization; x p (q) is the evaluation index before initialization; x1(q) is the first data of the evaluation index.

4. The method for intelligent prediction of coal mine gas outbursts integrating prior knowledge and segmented training according to claim 1 is characterized in that: In step S3, the feature extractor operates as follows: H sensitive =H original *T; Among them, H sensitive is the sensitive index feature matrix obtained after processing by the feature extractor, H original is a matrix with n rows and m columns, where n is the number of sample data, m is the dimension of sample data, and H original Each prediction index is defined as X1 to X9, X nm H original The elements in , T is the sensitive indicator feature extractor, and two sensitive indicators are selected as prior knowledge and placed in the second and third columns of the sensitive indicator feature extractor T.

5. The method for intelligent prediction of coal mine gas outburst integrating prior knowledge and segmented training according to claim 1 is characterized in that: The loss function of the segmented training MLP network model is as follows: Among them, Y i is the label of the first stage training sample; is the prediction result of the first stage training sample; Y j is the label of the second stage training sample; Prediction results for the second stage training samples; LOSS 全局 is the global loss; L 阶段1 is the loss in the first stage; L 阶段2 is the loss of the second stage; i and j are the indexes of the samples.

Images