A mine microseismic multivariate time series prediction method based on a double-flow structure
By using a multivariate time series prediction method for mine microseismic events based on a dual-flow structure, the problems of single feature extraction and poor generalization in coal mining microseismic signal prediction are solved. This method enables timely and accurate early warning of high-energy microseismic events, improving the prediction accuracy and efficiency of coal mine safety production.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- LIAONING UNIVERSITY
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-19
AI Technical Summary
Existing technologies for predicting microseismic signals in coal mining suffer from limitations such as single feature extraction and poor generalization, making it difficult to achieve timely and accurate early warning of high-energy microseismic events and affecting safe production in coal mines.
A multivariate time series prediction method for mine microseismic events based on a dual-flow structure is adopted. Through data preprocessing, feature extraction and fusion of variables and time dimensions, dynamic sparse graphs are constructed using graph convolution and global attention mechanisms to generate low-rank basis matrices and covariance patterns. Multi-dimensional information is fused, and finally, prediction results are generated through a linear projection layer.
It significantly improves the prediction accuracy of microseismic events, enables timely and accurate early warning of high-energy microseismic events, reduces model computational redundancy, adapts to the real-time monitoring needs of mine sites, reduces the risk of geological disasters, and ensures safe production.
Smart Images
Figure CN121679672B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of microseismic prediction in coal mining, specifically involving a method for predicting multiple microseismic signals in coal mining. Background Technology
[0002] Currently, microseismic signal prediction is an important means to ensure safe coal mining and improve mining efficiency. Accurate prediction and preventative measures are of significant practical importance in improving the safety of coal mining operations, reducing the probability of accidents, and minimizing economic losses. Microseismic events refer to the release of elastic energy caused by coal and rock fracturing during mining disturbances. High-energy microseismic events may induce coal mine rock bursts or other coal and rock dynamic disasters.
[0003] Therefore, predicting hazards in advance and taking corresponding measures to mitigate or avoid losses caused by dynamic disasters such as rockbursts is an urgent problem to be solved in the field of coal mine safety production. Real-time monitoring and analysis of microseismic events during coal mining, and research on the prediction of microseismic events, are of great significance for preventing and mitigating dynamic disasters in mines and ensuring the safety of coal mine production.
[0004] Microseismic monitoring data can be considered time-series data because it is ordered over time and exhibits a certain degree of time dependence. Time-series analysis can uncover inherent patterns in the data, accurately capture trends, and thus lay a foundation for predicting future events. By applying time-series forecasting methods, it is possible to effectively analyze and predict potential microseismic events during coal mining, thereby providing early warning information for mine management. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies, this invention provides a multivariate time series prediction method for mine microseismic events based on a dual-flow structure.
[0006] The technical solution of this invention is: a multivariate time series prediction method for mine microseismic events based on a dual-flow structure, the steps of which are as follows:
[0007] Step 1) Preprocess the microseismic monitoring data;
[0008] The specific method is as follows:
[0009] For input multivariate time series data Standardization and missing value handling are performed, where N is the number of variables and T is the time step, to adapt the data quality to the model input and eliminate the dimensional differences between data features. In order to eliminate dimensional differences, Z-Score normalization is used, as shown in formula (1):
[0010]
[0011] in: Let σ(n) be the mean of the nth variable, and let σ(n) be the standard deviation of the nth variable.
[0012] Step 2) Process the microseismic monitoring data at the variable level;
[0013] The specific method is as follows:
[0014] Step 2.1) Normalize the data Reshape by variable dimension The feature is mapped to a high-dimensional feature space through an embedding layer and then overlaid with time stamps, i.e., year, month, day, and hour encodings, as shown in the following formula:
[0015]
[0016] in: For the variable embedding layer, the parameter matrix is The time stamp vector is , dimension After aligning with the embedded features through linear projection, the final The initial characteristics of the nth variable;
[0017] Perform a Fast Fourier Transform on the input multivariate time-series data to map the time-domain signal to the frequency domain:
[0018]
[0019] By setting the hyperparameter top_k, the top k frequency points with the largest amplitude are selected, and the corresponding period is obtained accordingly. By normalizing different periods, the corresponding weights of the subsequent fusion map features are obtained.
[0020] Step 2.2) Based on the initial features of all variables, perform graph convolution calculation to dynamically correlate the variables and output intermediate features that fuse the correlation information;
[0021] Generate the required low-rank basis matrix The initialization uses a standard normal distribution, which maps the correlation features of the low-rank space back to the original N-dimensional variable space, reducing redundant calculations.
[0022] Generate covariance mode matrix M represents the number of covariance patterns, with each row corresponding to a typical variable association pattern; based on the preprocessed variable features and time stamps, dynamic weights that change with time steps are generated.
[0023]
[0024] in, The similarity between the query vector and the key vector is calculated. A higher similarity results in a higher score and a stronger correlation between the variables. Based on this score, the original features and dynamically associated features are fused and transformed into dimensions suitable for subsequent prediction layers. The specific formula is as follows:
[0025]
[0026] in, The contribution ratio of original features to associated features is controlled and adaptively optimized during model training. This is achieved by controlling... The magnitude of the variable is used to obtain different fusion characteristics between variables.
[0027] Step 3) Process the microseismic monitoring data in the time dimension;
[0028] The specific method is as follows:
[0029] Focusing on the changing patterns within the current time series itself, and processing... Divided along the time dimension Each non-overlapping patch contains [number] patches. Each time step, the formula is as follows:
[0030]
[0031] Calculate the mean along the variable dimension and integrate the time coupling characteristics of multiple variables to eliminate variable redundancy;
[0032]
[0033] A dynamic sparse graph within the sequence is constructed, and the Pearson correlation coefficient is used to analyze the similarity between different patch blocks. The higher the Pearson correlation coefficient, the stronger the correlation between patches. Based on the calculated similarity as weights, the top k most relevant patches are selected as adjacent nodes to construct an intra-sequence adjacency matrix. The specific formula is as follows:
[0034]
[0035] in, This represents the Pearson correlation coefficient between patch_i and patch_j;
[0036] A global attention mechanism within a sequence is introduced, which weights and aggregates information by calculating the correlation between each patch. The specific formula is shown below:
[0037]
[0038] in, This represents the update feature of the i-th patch. The relation coefficient determines the amount of information each feature receives from other patches.
[0039] Step 4) Concatenate the prediction results of the variable module and the time module to obtain the prediction result.
[0040] The predictions from the time dimension and the predictions from the variable dimension are fused together, and these two features are concatenated along the last dimension to obtain the latest feature representation. The formula is shown below:
[0041]
[0042] A linear projection layer is introduced, using a trainable linear mapping matrix. The concatenated features are mapped to the output space to obtain the final prediction result.
[0043]
[0044] The beneficial effects of this invention are as follows: It can extract the temporal and spatial correlation features of microseismic signals separately, achieving deep fusion of multi-dimensional information and significantly improving prediction accuracy. It can efficiently process multi-variable coupled data such as microseismic energy, frequency, and amplitude, solving the problems of single feature extraction and poor generalization in traditional methods, resulting in more timely and accurate early warnings for high-energy microseismic events. Simultaneously, it reduces model computational redundancy, adapts to the real-time monitoring needs of mine sites, provides reliable safety prediction basis for mining operations, effectively reduces geological disaster risks, ensures personnel and equipment safety, and improves the safety and production efficiency of mining operations. Attached Figure Description
[0045] Figure 1 The following is a system flowchart of a multivariate time series prediction method for mine microseismic events based on a dual-flow structure.
[0046] Figure 2 : Schematic diagram of the overall architecture of this invention.
[0047] Figure 3 This invention generates an adaptive spatial structure schematic diagram. Detailed Implementation
[0048] A multivariate time series prediction method for mine microseismic events based on a dual-flow structure, comprising the following steps:
[0049] Step 1) Preprocess the microseismic monitoring data;
[0050] Step 2) Process the microseismic monitoring data in the variable module;
[0051] Step 3) Process the microseismic monitoring data in the time module;
[0052] Step 4) Merge the prediction results from the variable module and the time module to obtain the prediction result.
[0053] In step 1), N key sensors are selected, and data collected includes the energy (J) of the microseismic event, frequency (times / hour), and source coordinates. The sampling time interval is set to 1 hour. Since signal transmission failures may lead to data loss, we use linear interpolation to fill in the missing data, as shown in the following formula:
[0054]
[0055] The collected data is named (N is the number of variables, T is the time step)
[0056] Normalizing the data to a distribution with a mean of 0 and a variance of 1 ensures that the data quality is suitable for the model input. This process can eliminate dimensional differences between data features, improve model training stability, and enhance convergence efficiency.
[0057]
[0058] In order to eliminate dimensional differences, Z-Score normalization is used. Let σ(n) be the mean of the nth variable, and σ(n) be the standard deviation of the nth variable. We retain... σ(n) is used for the inverse normalization of subsequent prediction results.
[0059] In step 2), we focus on feature fusion between different variables. First, we normalize the data. Reshape by variable dimension The feature is mapped to a high-dimensional feature space through an embedding layer and then overlaid with time stamps (year, month, day, and hour codes), as shown in the following formula:
[0060]
[0061] Among them: Among them, For the variable embedding layer, the parameter matrix is The time stamp vector is , dimension After aligning with the embedded features through linear projection, the final Let be the initial characteristics of the nth variable.
[0062] We perform a Fast Fourier Transform on the input multivariate time-series data to map the time-domain signal to the frequency domain:
[0063]
[0064] By setting the hyperparameter top_k, we can filter out the top k frequency points with the largest amplitude and obtain the corresponding period. By normalizing different periods, we can obtain the corresponding weights of the features in the subsequent fusion map.
[0065] Based on the initial features of all variables, we implement graph convolution computation for dynamic relationships between variables, outputting intermediate features that fuse the relationship information. First, we generate the required low-rank basis matrix. The initialization uses a standard normal distribution, which maps the correlation features in the low-rank space back to the original N-dimensional variable space, reducing redundant computation. Then, the covariance pattern matrix is generated. M represents the number of covariance patterns, with each row corresponding to a typical variable association pattern.
[0066] Based on the preprocessed variable features and time stamps, dynamic weights that change with time steps are generated.
[0067]
[0068] We calculate the similarity between the query vector and the key vector. The higher the similarity, the higher the score, and the stronger the correlation between the variables. Based on this score, we fuse the original features and dynamically associated features, and convert them into dimensions suitable for subsequent prediction layers. The specific formula is as follows:
[0069]
[0070] The contribution ratio between original features and associated features is controlled and adaptively optimized during model training. For each covariance pattern, the scores obtained in (6) are weighted and summed to obtain the correlation features of temporal dynamics. The low-rank basis matrix U is then used to further refine the dynamic correlation features. Mapping back to the original variable space, we get Therefore, we can obtain the fusion characteristics between variables.
[0071] In step 3), we focus on the changing patterns within the current time series itself. First, we process the... Divided along the time dimension Each non-overlapping patch contains [number] patches. Each time step, the formula is as follows:
[0072]
[0073] Subsequently, the mean is calculated along the variable dimension, and the time coupling characteristics of multiple variables are integrated to eliminate variable redundancy.
[0074]
[0075] Next, we construct an intra-sequence dynamic sparse graph. We use the Pearson correlation coefficient to analyze the similarity between different patch blocks; a higher Pearson correlation coefficient indicates a stronger correlation between patches. Then, based on the calculated similarity as weights, we select the k most relevant patches as neighboring nodes to construct the intra-sequence adjacency matrix, as shown in the following formula:
[0076]
[0077] in, This represents the Pearson correlation coefficient between patch_i and patch_j. This is the adjacency matrix that is constructed.
[0078] While adjacency matrices can progressively aggregate local features between different patches, this local dependency modeling is insufficient for capturing global information across the entire time series. Therefore, we introduce an intra-sequence global attention mechanism that weights and aggregates information by calculating the correlations between each patch. The specific formula is shown below:
[0079]
[0080] in, This represents the update feature of the i-th patch. The relation coefficient determines the amount of information each feature receives from other patches. This is the scaled attention score. These are all learnable parameter matrices, optimized through backpropagation.
[0081] In step 4), we fuse the predictions from the time dimension with the predictions from the variable dimension. We concatenate these two features along the last dimension to obtain the latest feature representation. The formula is shown below:
[0082]
[0083] Then we introduce a linear projection layer, using a trainable linear mapping matrix. The concatenated features are mapped to the output space.
[0084]
[0085] To obtain the final prediction result .
Claims
1. A mine microseismic multivariate time series prediction method based on a double-flow structure, characterized in that, The steps are as follows: Step 1) Preprocess the microseismic monitoring data; In step 1), the specific method is as follows: For input multivariate time series data Standardization and missing value handling are performed, where N is the number of variables and T is the time step, to adapt the data quality to the model input and eliminate the dimensional differences between data features. In order to eliminate dimensional differences, Z-Score normalization is used, as shown in formula (1): wherein: is the mean of the nth variable and σ(n) is the standard deviation of the nth variable; Step 2) Process the microseismic monitoring data at the variable level; In step 2), the specific method is as follows: Step 2.1) Normalize the data Reshape by variable dimension The feature is mapped to a high-dimensional feature space through an embedding layer and then overlaid with time stamps, i.e., year, month, day, and hour encodings, as shown in the following formula: in: For the variable embedding layer, the parameter matrix is The time stamp vector is , dimension After aligning with the embedded features through linear projection, the final The initial characteristics of the nth variable; Perform a Fast Fourier Transform on the input multivariate time-series data to map the time-domain signal to the frequency domain: By setting the hyperparameter top_k, the top k frequency points with the largest amplitude are selected, and the corresponding period is obtained accordingly. By normalizing different periods, the corresponding weights of the subsequent fusion map features are obtained. Step 2.2) Based on the initial features of all variables, perform graph convolution calculation to dynamically correlate the variables and output intermediate features that fuse the correlation information; Generate the required low-rank basis matrix The initialization uses a standard normal distribution, which maps the correlation features of the low-rank space back to the original N-dimensional variable space, reducing redundant calculations. Generating a covariance pattern matrix M is the number of covariance patterns, each row corresponds to a typical variable correlation pattern; based on the pre-processed variable characteristics and time markers, continue to generate dynamic weights that change with time steps; wherein, The similarity of the query vector and the key vector is calculated, the higher the similarity, the higher the score, the stronger the correlation between variables, the original features and dynamic correlation features are fused according to the score, and are converted into a dimension suitable for subsequent prediction layers; the specific formula is as follows: in, The contribution ratio of original features to associated features is controlled and adaptively optimized during model training. This is achieved by controlling... The magnitude of the variables is used to obtain different fusion features between them; Step 3) Process the microseismic monitoring data in the time dimension; Step 4) Concatenate the prediction results of the variable module and the time module to obtain the prediction result.
2. The mine microseismic multivariate time series prediction method based on a double-flow structure according to claim 1, characterized in that, In step 3), the specific method is as follows: Focusing on the changing patterns within the current time series itself, and processing... Divided along the time dimension Each non-overlapping patch contains [number] patches. Each time step, the formula is as follows: Calculate the mean along the variable dimension and integrate the time coupling characteristics of multiple variables to eliminate variable redundancy; A dynamic sparse graph within the sequence is constructed, and the Pearson correlation coefficient is used to analyze the similarity between different patch blocks. The higher the Pearson correlation coefficient, the stronger the correlation between patches. Based on the calculated similarity as weights, the top k most relevant patches are selected as adjacent nodes to construct an intra-sequence adjacency matrix. The specific formula is as follows: wherein, represents the Pearson correlation coefficient between patch_i and patch_j; A global attention mechanism within a sequence is introduced, which weights and aggregates information by calculating the correlation between each patch. The specific formula is shown below: wherein, represents the updated feature of the i-th patch, is a relation coefficient that determines the amount of information each feature receives from other patches.
3. The mine microseismic multivariate time series prediction method based on a double-flow structure according to claim 1, characterized in that, In step 4), the specific method is as follows: The prediction result from the time dimension is fused with the prediction result from the variable dimension, and the two features are spliced along the last dimension to obtain the latest feature representation , the formula is as follows: A linear projection layer is introduced, using a trainable linear mapping matrix. The concatenated features are mapped to the output space to obtain the final prediction result. 。