Rock burst multi-index composite evaluation early warning method based on informer model
By using a multi-index composite evaluation warning method based on the informer model in impact ground pressure prediction, and using the sparse self-attention and self-attention distillation mechanism, the problems of low prediction accuracy and large calculation amount in the prior art are solved, and more accurate and real-time impact ground pressure warning is achieved.
Patent Information
- Application Number
- CN202411758367.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-03
- Publication Date
- 2025-03-04
AI Technical Summary
The prior art has problems in the prediction of impact ground pressure, low prediction accuracy, large calculation amount, and unreal-time early warning evaluation.
The impact ground pressure multi-index composite evaluation early warning method based on the informer model is adopted. Through field data collection and processing, the sparse self-attention mechanism and self-attention distillation mechanism are used to predict it, and the multi-index composite evaluation is performed based on the ground sound energy, micro-seismic frequency and pressure data on the hydraulic pillar are combined.
The accuracy and real-time nature of impact ground pressure warning is improved, the calculation complexity is reduced, and the warning results are made more accurate and reasonable through a multi-index composite evaluation system.
Smart Images
Figure CN119558542A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of coal mine safety technology, and in particular to coal mine rock burst early warning technology. Background Art
[0002] With the continuous development of the economy and society, the demand for natural resources has also increased, and mineral resources are one of them. It can be said that mineral resources are the basic raw materials for the survival and development of human society, and they play a vital role in the development of the entire national economy. At present, the development and utilization of mineral resources is in the stage of "from shallow to deep, from easy to difficult". Deep resources have huge potential, and the development and utilization of deep resources is inevitable for the development of the mining industry. However, deep mineral mining faces the situation of "three highs and one disturbance", namely "high ground stress, high permeability ground pressure, high ground temperature and strong mining disturbance", which makes the environment of deep mineral mining very harsh, and it is very easy to cause various engineering geological disasters, rock burst is one of them. Therefore, early warning of rock burst is a key step in safe underground mining.
[0003] In the past, most of the early warnings for rock burst were measured using mechanical devices. Although the results were highly accurate, they consumed a lot of manpower and material resources, and could not effectively predict rock burst in real time. In addition, the prediction methods based on transformer and LSTM models have a large amount of calculation and are not particularly accurate for long-term predictions. Summary of the invention
[0004] In view of this, the purpose of the present invention is to provide a multi-index composite evaluation and early warning method for rock burst based on an informer model to solve the technical problems of rock burst prediction in terms of prediction accuracy, prediction calculation amount, early warning evaluation, etc.
[0005] The multi-index composite evaluation and early warning method for rock burst based on the informer model of the present invention includes:
[0006] 1) Collecting geosound energy monitoring data, microseismic frequency monitoring data and hydraulic support pressure data with the same time series on site, arranging and combining the collected data according to the time series to obtain sample data sets of the three types of data, and dividing the sample data sets into training set, test set and validation set according to a certain ratio;
[0007] II) Use the training set, validation set, and test set to train, test, and validate the informer model respectively;
[0008] Training the informer model includes:
[0009] 1) The sample data set is embedded and positionally encoded to obtain the initial input data of the encoding part of the informer model;
[0010] The encoding part of the informer model performs three linear transformations on the initial input data to obtain three matrices Q, K, and V. The sparse self-attention operation is performed on the three matrices Q, K, and V to obtain the attention matrix. The self-attention distillation operation is performed on the attention matrix to obtain the final feature map.
[0011] 2) The sample data set is subjected to embedding operation and position encoding splicing to obtain the first half of the input data of the decoder part of the informer model, and the feature map obtained in step 1) is used as the second half of the input data of the decoder part of the informer model; the decoder part of the informer model performs a masked multi-head sparse self-attention operation on the first half of the input data, and then the decoder part of the informer model performs a multi-head attention operation on the second half of the data and the data output by the masked multi-head sparse self-attention operation to obtain the output of the decoder part of the informer model;
[0012] 3) The output of the decoded part of the informer model is input into a fully connected layer, which outputs a series of prediction results;
[0013] III) using the informer model obtained through training, testing and verification in step II) to process the real-time collected geosound energy monitoring data, microseismic frequency monitoring data and hydraulic support pressure data to obtain a prediction value;
[0014] IV) referring to the graded warning indicators of mines with similar geological and mining conditions in the past, determine the energy index range and intensity ratio range of the predicted data obtained in step III), and finally determine the time and warning intensity of the predicted ground pressure impact event.
[0015] Further, in step IV) the following is included:
[0016] Determine the ratio R1 of the predicted pressure data of the hydraulic support obtained in step III) to the strength of the ore rock, and divide the basic warning level T1 according to the value of R1, as follows:
[0017]
[0018] T1 is 1, 2, 3, and 4, which represent no rock burst danger, weak danger, medium danger, and strong danger, respectively;
[0019] Based on the geosound energy prediction data obtained in step III), the predicted evolution trend of geosound energy is depicted, which is generally divided into three trends: (1) basically unchanged; (2) slowly rising; (3) rapidly rising; these three trends are assigned -1, 0, and 1 values to the correction warning level T2 in turn, as shown below:
[0020]
[0021] Based on the microseismic frequency monitoring data obtained in step III), the predicted evolution trend of the microseismic frequency is depicted, which can be generally divided into two trends: (1) basically maintained at 0; (2) abnormally increased; these two trends are assigned the values of 0 and 1 of the supplementary warning level T3 in turn, as shown below:
[0022]
[0023] Add the obtained T1, T2 and T3 values to obtain the final warning level T and take corresponding measures.
[0024] Furthermore, the embedding operation is to increase the dimension of the sample data set.
[0025] Furthermore, the position code concatenation includes: using sine and cosine functions of different frequencies to generate the position code, and then adding it to the feature vector of the corresponding position; the dimension of the position vector is consistent with the dimension of the feature vector; the position code generation formula is:
[0026]
[0027] In the formula, PE represents the position code; pos represents the position of the current character in the sequence; d model Represents the dimension of the current character embedding.
[0028] Further, the sparse self-attention operation includes:
[0029] (1) Randomly sample n k for each K matrix, use the n k to do the inner product with each q of the Q matrix, and obtain the correlation between these n k and each q; thus calculate the sparsity score M of each q; the calculation formula of the sparsity score M is as follows:
[0030]
[0031] Where: q i is the vector value of the i-th row in the Q matrix; is the transpose of the vector value of the i-th row in the K matrix; represents the normalization operation; L k It represents the length of K matrix;
[0032] (2) Select n q with the highest sparsity score M, and then calculate the inner product of these n q for each k of the K matrix to obtain the correlation results of these n q;
[0033] (3) The remaining q are not calculated, and the average value of the n q correlation results in step (2) is directly taken. This ensures that the final QK matrix can match the size of the V matrix, and the final attention matrix is calculated.
[0034] Furthermore, the self-attention distillation operation includes: performing several convolution pooling operations on the attention matrix, pruning the input dimension, and after each convolution pooling operation, the length of the input sequence is halved; after each convolution pooling operation, the attention matrix of the sparse self-attention mechanism needs to be calculated on the processed data, and finally the final feature map is obtained; therefore, the self-attention distillation process from the i-th layer to the i+1-th layer is:
[0035]
[0036] Where: MaxPool1d is the maximum pooling operation; ELU is the activation function of the one-dimensional convolution operation of Conv1d; Conv1d represents the one-dimensional convolution operation.
[0037] Beneficial effects of the present invention:
[0038] 1. The rock burst multi-index composite evaluation and early warning method based on the informer model of the present invention adopts two kinds of field data for prediction and processing, so that the rock burst early warning is more accurate.
[0039] 2. The multi-index composite evaluation and early warning method for rock burst based on the informer model of the present invention adopts the informer model. Compared with the previous prediction model, the calculation complexity of the self-attention matrix and related parameters is greatly reduced through the sparse self-attention mechanism and the self-attention distillation mechanism. In addition, a one-step decoding method is also adopted, and one-step decoding can obtain subsequent multiple prediction data. Therefore, the present invention can better ensure the real-time nature of the prediction and ensure the accuracy of long-term prediction.
[0040] 3. The rock burst multi-index composite evaluation and early warning method based on the informer model of the present invention adopts a rock burst early warning evaluation system based on ground sound energy monitoring data, microseismic frequency monitoring data and hydraulic support pressure data prediction data, and gives measures that should be taken for each evaluation. Combining the prediction results of the three makes the prediction results more accurate and the early warning evaluation more reasonable. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 This is the overall flow chart of the rock burst warning method based on the informer model.
[0042] Figure 2 This is the overall architecture diagram of the informer prediction method.
[0043] Figure 3 This is a diagram of the evaluation process for rock burst warning and the corresponding measures to be taken. DETAILED DESCRIPTION
[0044] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0045] The multi-index composite evaluation and early warning method for rock burst based on the informer model in this embodiment includes:
[0046] I) Collect geosound energy monitoring data, microseismic frequency monitoring data and hydraulic support pressure data with the same time series on site, arrange and combine the collected data according to the time series to obtain sample data sets of the three types of data, divide the sample data sets into training set, test set and validation set according to a certain ratio, and predict geosound monitoring data and hydraulic support pressure data respectively. The training set is used to train the informer model to obtain the training model; the test set is used to evaluate the performance of the training model; the validation set is used to adjust and select the final model network parameters.
[0047] II) Use the training set, validation set, and test set to train, test, and validate the informer model respectively;
[0048] Training the informer model includes:
[0049] 1) The sample data set is embedded and positionally encoded to obtain the initial input data of the encoding part of the informer model.
[0050] The embedding operation is to increase the dimension of the sample data set and amplify the features of the sample data set.
[0051] The position code splicing includes: using sine and cosine functions of different frequencies to generate position codes, and then adding them to the feature vectors of the corresponding positions; the dimension of the position vector is consistent with the dimension of the feature vector; the position code generation formula is:
[0052]
[0053]
[0054] In the formula, PE represents the position code; pos represents the position of the current character in the sequence; d model Represents the dimension of the current character embedding.
[0055] The encoding part of the informer model performs three linear transformations on the initial input data, namely, Q , W K , W V The weight matrix is multiplied to obtain three matrices Q, K, and V. It should be noted that W Q , W K , W VThe matrix is constantly changing during the model training phase.
[0056] The sparse self-attention operation is performed on the three matrices Q, K, and V to obtain the attention matrix, and the self-attention distillation operation is performed on the attention matrix to obtain the final feature map.
[0057] The sparse self-attention operation includes:
[0058] (1) Randomly sample n k for each K matrix, use the n k to do the inner product with each q of the Q matrix, and obtain the correlation between these n k and each q; thus calculate the sparsity score M of each q; the calculation formula of the sparsity score M is as follows:
[0059]
[0060] Where: q i is the vector value of the i-th row in the Q matrix; is the transpose of the vector value of the i-th row in the K matrix; represents the normalization operation; L k It represents the length of K matrix;
[0061] (2) Select n q with the highest sparsity score M, and then calculate the inner product of these n q for each k of the K matrix to obtain the correlation results of these n q;
[0062] (3) The remaining q are not calculated, and the average value of the n q correlation results in step (2) is directly taken. This ensures that the final QK matrix can match the size of the V matrix, and the final attention matrix is calculated.
[0063] The self-attention distillation operation includes: performing several convolution pooling operations on the attention matrix, pruning the input dimension, and halving the input sequence length after each convolution pooling operation; after each convolution pooling operation, the attention matrix of the sparse self-attention mechanism needs to be calculated on the processed data, and finally the final feature map is obtained. Therefore, the self-attention distillation process from the i-th layer to the i+1-th layer is:
[0064]
[0065] Where: MaxPool1d is the maximum pooling operation; ELU is the activation function of the one-dimensional convolution operation of Conv1d; Conv1d represents the one-dimensional convolution operation.
[0066] 2) The second half of the input data after the embedding operation and position encoding in step 1) is used as the first half of the input data of the decoder part of the informer model to guide the prediction of the informer model. The second half of the input data of the decoder part of the informer model is a mask matrix, that is, a 0 matrix with the same shape as the predicted target. The decoder part of the informer model performs a masked multi-head sparse self-attention operation on the input data, and then performs a multi-head attention operation on the data output by the masked multi-head sparse self-attention operation to obtain the output of the decoder part of the informer model; the masked multi-head sparse self-attention operation is basically the same as the sparse self-attention operation of the decoding part, and the difference is that the inner product of the mask part in the input data is set to negative infinity to shield the attention of the front part to the mask part, so that it has no effect when the parameters are updated. The multi-head attention operation is also the inner product operation of the three matrices Q, K, and V. Unlike the sparse self-attention operation, the multi-head attention operation does not screen the Q matrix. In addition, the K and V matrices come from the final output of the encoding part, and the Q matrix comes from the output of the masked multi-head sparse self-attention operation. After the decoding part, there is a vector for each position to be predicted.
[0067] 3) The final output of the decoded part of the informer model is input into a fully connected layer, which outputs a series of prediction results.
[0068] III) using the optimal informer model obtained through training, testing and verification in step II) to process the real-time collected geosound energy monitoring data, microseismic frequency monitoring data and hydraulic support pressure data to obtain a prediction value;
[0069] IV) Referring to the graded warning indicators of mines with similar geological and mining conditions in the past, determine the ratio R1 of the pressure prediction data of the hydraulic support obtained in step III) to the strength of the ore rock, and divide the basic warning level T1 according to the value of R1, which is specifically divided as follows:
[0070]
[0071] Among them, T1 is 1, 2, 3, and 4, respectively representing no impact ground pressure danger, weak danger, medium danger, and strong danger.
[0072] Based on the geosound energy prediction data obtained in step III), the predicted evolution trend of geosound energy is depicted, which is generally divided into three trends: (1) basically unchanged; (2) slowly rising; (3) rapidly rising. These three trends are assigned -1, 0, and 1 values to the correction warning level T2, as shown below:
[0073]
[0074] Based on the microseismic frequency monitoring data obtained in step III), the predicted evolution trend of the microseismic frequency is depicted, which is generally divided into two trends: (1) basically maintained at 0; (2) abnormally increased. These two trends are assigned the values of 0 and 1 of the supplementary warning level T3 in turn, as shown below:
[0075]
[0076] Add the obtained T1, T2 and T3 values to get the final warning level T. Therefore, the value range of T is between 0-6, and it is divided into 4 danger levels, and corresponding measures are taken. The details are as follows: when T is 0 or 1, there is no possibility of rock burst, and normal production is carried out; when T is 2, there is no possibility of rock burst, and the monitoring of rock burst at the data collection site needs to be strengthened; when T is 3 or 4, weak rock burst is possible, and rock burst management measures need to be taken to reduce the intensity of rock burst and the size of the harm caused in the future; when T is 5 or 6, production should be stopped and wait for the rock burst event to occur and subside.
[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A multi-index composite evaluation and early warning method for rock burst based on an informer model, characterized by: include: 1) Collecting geosound energy monitoring data, microseismic frequency monitoring data and hydraulic support pressure data with the same time series on site, arranging and combining the collected data according to the time series to obtain sample data sets of the three types of data, and dividing the sample data sets into training set, test set and validation set according to a certain ratio; II) Use the training set, validation set, and test set to train, test, and validate the informer model respectively; Training the informer model includes: 1) The sample data set is embedded and positionally encoded to obtain the initial input data of the encoding part of the informer model; The encoding part of the informer model performs three linear transformations on the initial input data to obtain three matrices Q, K, and V. The sparse self-attention operation is performed on the three matrices Q, K, and V to obtain the attention matrix. The self-attention distillation operation is performed on the attention matrix to obtain the final feature map. 2) The sample data set is subjected to embedding operation and position encoding splicing to obtain the first half of the input data of the decoder part of the informer model, and the feature map obtained in step 1) is used as the second half of the input data of the decoder part of the informer model; the decoder part of the informer model performs a masked multi-head sparse self-attention operation on the first half of the input data, and then the decoder part of the informer model performs a multi-head attention operation on the second half of the data and the data output by the masked multi-head sparse self-attention operation to obtain the output of the decoder part of the informer model; 3) The output of the decoded part of the informer model is input into a fully connected layer, which outputs a series of prediction results; III) using the informer model obtained through training, testing and verification in step II) to process the real-time collected geosound energy monitoring data, microseismic frequency monitoring data and hydraulic support pressure data to obtain a prediction value; IV) referring to the graded warning indicators of mines with similar geological and mining conditions in the past, determine the energy index range and intensity ratio range of the predicted data obtained in step III), and finally determine the time and warning intensity of the predicted ground pressure impact event.
2. The multi-index composite evaluation and early warning method for rock burst based on the informer model according to claim 1 is characterized by: In step IV) it includes: Determine the ratio R1 of the predicted pressure data of the hydraulic support obtained in step III) to the strength of the ore rock, and divide the basic warning level T1 according to the value of R1, as follows: T1 is 1, 2, 3, and 4, which represent no rock burst danger, weak danger, medium danger, and strong danger, respectively; Based on the geosound energy prediction data obtained in step III), the predicted evolution trend of geosound energy is depicted, which is generally divided into three trends: (1) basically unchanged; (2) slowly rising; (3) rapidly rising; these three trends are assigned -1, 0, and 1 values to the correction warning level T2 in turn, as shown below: Based on the microseismic frequency monitoring data obtained in step III), the predicted evolution trend of the microseismic frequency is depicted, which can be generally divided into two trends: (1) basically maintained at 0; (2) abnormally increased; these two trends are assigned the values of 0 and 1 of the supplementary warning level T3 in turn, as shown below: Add the obtained T1, T2 and T3 values to obtain the final warning level T and take corresponding measures.
3. The multi-index composite evaluation and early warning method for rock burst based on the informer model according to claim 1 is characterized by: The embedding operation is to increase the dimension of the sample data set.
4. The multi-index composite evaluation and early warning method for rock burst based on the informer model according to claim 1 is characterized by: The position code splicing includes: using sine and cosine functions of different frequencies to generate position codes, and then adding them to the feature vectors of the corresponding positions; the dimension of the position vector is consistent with the dimension of the feature vector; the position code generation formula is: In the formula, PE represents the position code; pos represents the position of the current character in the sequence; d model Represents the dimension of the current character embedding.
5. The multi-index composite evaluation and early warning method for rock burst based on the informer model according to claim 1 is characterized by: The sparse self-attention operation includes: (1) Randomly sample n k for each K matrix, use the n k to do the inner product with each q of the Q matrix, and obtain the correlation between these n k and each q; thus calculate the sparsity score M of each q; the calculation formula of the sparsity score M is as follows: Where: q i is the vector value of the i-th row in the Q matrix; is the transpose of the vector value of the j-th row in the K matrix; represents the normalization operation; L k It represents the length of K matrix; (2) Select n q with the highest sparsity score M, and then calculate the inner product of these n q with each k of the K matrix to obtain the correlation results of these n q; (3) The remaining q are not calculated, and the average of the n q correlation results in step (2) is directly taken. This ensures that the final QK matrix can match the size of the V matrix, and the final attention matrix is calculated.
6. The multi-index composite evaluation and early warning method for rock burst based on the informer model according to claim 5 is characterized by: The self-attention distillation operation includes: performing several convolution pooling operations on the attention matrix, pruning the input dimension, and halving the input sequence length after each convolution pooling operation; after each convolution pooling operation, the attention matrix of the sparse self-attention mechanism needs to be calculated on the processed data, and finally the final feature map is obtained; in addition, the self-attention distillation process from the i-th layer to the i+1-th layer is: Where: MaxPool1d is the maximum pooling operation; ELU is the activation function of the one-dimensional convolution operation of Conv1d; Conv1d represents the one-dimensional convolution operation.