Micro-seismic intelligent positioning method based on deep feedforward neural network

Abstract

The invention discloses a micro-seismic intelligent positioning method based on a deep feed-forward neural network, which combines a multi-head self-attention mechanism and a deep feed-forward neural network model of a residual module to carry out micro-seismic event positioning, and comprises the following steps: firstly, collecting spatio-temporal information of a plurality of monitoring points as input data, calculating time difference and recombining the time difference into a feature matrix; the dimension difference is eliminated after standardization processing; subspace splitting is carried out on the spatio-temporal information by using a multi-head self-attention mechanism, so that a self-attention weight is distributed in each subspace, and relevance of different feature dimensions is obtained; on the basis, gradient disappearance is prevented through a residual module, then a spatial difference time rule in layered wave velocity model data is extracted, a deep feedforward neural network model is trained, and the training stability of the deep feedforward neural network model is ensured; and finally, the model output layer outputs predicted three-dimensional coordinates of the seismic source. The method has the advantages of high positioning precision, high calculation efficiency and the like, and is particularly suitable for real-time monitoring and positioning of micro-seismic events in a layered wave velocity model complex geological environment.

Description

Technical Field

[0001] The present invention relates to a microseismic event locating method, in particular to a microseismic intelligent locating method based on a deep feedforward neural network. Background Art

[0002] When micro-defects within a rock mass are damaged by surrounding stresses, they generate weak seismic waves, known as microseismic waves. Microseismic monitoring technology deploys sensors within a target area to collect and process microseismic signals generated by rock fractures, enabling real-time dynamic monitoring of the fractured area.

[0003] Microseismic event location is a core part of microseismic monitoring and is of great significance for revealing the dynamic evolution of underground rock stress fields. Microseismic events and earthquakes share certain similarities in their focal mechanisms, so relevant seismic data processing methods provide a reference for achieving microseismic source location. Depending on the required information, traditional microseismic location methods can be divided into travel-time-based location methods and waveform-based location methods. From the perspective of method principles and computational efficiency, travel-time-based location methods are simpler and more efficient than waveform-based location methods. However, neither method can take into account the nonlinear characteristics of the true wave velocity model, that is, the accuracy of microseismic source location is low in complex geological environments such as layered wave velocity models.

[0004] In view of the limitation that traditional positioning methods are only applicable to a single wave velocity model, how to provide a new method that can take into account the nonlinear characteristics under the real wave velocity model to achieve high-precision positioning in complex geological environments is the research direction required by this invention. Summary of the Invention

[0005] In response to the problems existing in the above-mentioned prior art, the present invention provides a microseismic intelligent positioning method based on a deep feedforward neural network. The deep feedforward neural network model combining a multi-head self-attention mechanism and a residual module can take into account the nonlinear characteristics under the real wave velocity model, thereby achieving high-precision positioning in complex wave velocity environments.

[0006] To achieve the above objectives, the present invention adopts a technical solution: a microseismic intelligent positioning method based on a deep feedforward neural network, comprising the following steps: S1. Select the spatiotemporal information of multiple monitoring points (including the three-dimensional coordinates of each monitoring point and the arrival time of each received earthquake signal) as input sample data; and simultaneously obtain the actual earthquake source location as verification sample data; S2. Calculate the arrival time difference of earthquake signals received between monitoring points , replace the original monitoring arrival data and form a one-dimensional array; S3, reorganize the one-dimensional array of step S2 into a two-dimensional feature matrix. Considering that the coordinate dimension and time dimension of the monitoring point are m and s respectively, the numerical scale difference between the two is large. In order to avoid the problem of excessive dependence of the network model on a certain feature due to numerical imbalance, the mean is used. and standard deviation , standardize different features and balance the values ​​of input data; S4. Flatten the normalized feature matrix through the Flatten layer of the deep feedforward neural network to adapt it to the receiving form of the hidden layer, and use the hidden layer to convert the data into a feature vector with higher discrimination; S5: Convert the more discriminative feature vectors from step S4 into feature matrices and pass them along two branches: the first branch extracts global dependencies through a multi-head self-attention mechanism and processes them through a normalization layer; the second branch acts as a skip connection and retains the original feature matrix. S6. The outputs of the two branches are merged in the concatenation layer of the deep feedforward neural network and deeply fused through the hidden layer. At the same time, in order to prevent the model from overfitting due to the excessive number of neurons, the Dropout module is added in this process to shield the outputs of some neurons; S7, passing the fusion result of step S6 to the residual module, and extracting the spatial time difference regularity through the stacked two layers of residual modules; S8. After training the deep feedforward neural network model according to the spatial time difference rule extracted in step S7, its output layer outputs the predicted three-dimensional coordinates of the earthquake source, and calculates the loss based on the verification sample data in step S1 to optimize the deep feedforward neural network model; S9. When the latest earthquake signals received by multiple monitoring points in step S1 are obtained, they are input into the model optimized in step S8, and the model outputs the three-dimensional coordinates of the earthquake source corresponding to the current earthquake signal.

[0007] Furthermore, the number of monitoring points selected in step S1 is four. The present invention only requires the spatiotemporal information of four monitoring points to accurately locate the microseismic source, which uses fewer monitoring points than other existing methods, facilitating data collection and processing.

[0008] Furthermore, the two-dimensional feature matrix in step S3 is specifically: Using the mean and standard deviation , standardize different features, specifically: Where, is the reorganized two-dimensional feature matrix; For the The monitoring point corresponds to the The standard value of a feature.

[0009] Furthermore, during the processing of step S4, the coordinate and time information in the data are in an associated space, and combined with the ReLU activation function, the mapping relationship between data features can be captured; coordinate and time features with smaller differences can be mapped to similar vectors, and vice versa, they can be mapped to vectors with more significant differences, thereby improving the model's ability to recognize different spatial and temporal difference features in subsequent work.

[0010] Furthermore, the multi-head self-attention mechanism in step S5 adjusts the dimension of the input matrix on top of the self-attention mechanism, thereby changing the structural dimension of the query matrix Q, the key matrix K and the value matrix V. Specifically, self-attention splits the input feature matrix into multiple subspaces with the same number of monitoring points, and each subspace focuses on different types of associated data, corresponding to the three directional coordinates and the monitored time difference between each monitoring point; then, through 、 、 Perform weighted summation in each subspace, assign attention weights, and obtain the splicing results of each subspace, so as to extract the global dependency relationship of features from multiple angles. The specific expression is: Where, The subspace obtained by splitting the input feature matrix; is the output of the multi-head self-attention layer; 、 and They are 、 and The corresponding weight.

[0011] Furthermore, in step S7, the residual module has two layers, and the specific calculation process of each layer of the residual module is: Where, and represent the input and output of the residual block respectively.

[0012] Furthermore, in step S8, the loss is calculated to optimize the deep feedforward neural network model. The specific optimization process is as follows: the verification sample data is divided into a training set and a test set in a ratio of 8:2, and the network model uses the trainingOptions function built into MATLAB to configure hyperparameters; through multiple rounds of hyperparameter adjustment, a reasonable mini-batch size (MiniBatchSize) and maximum training cycle (MaxEpochs) are set to optimize the memory usage of the model; at the same time, the initial learning rate is set to 0.0005, and the learning rate is halved every 5 data traversals to shorten the training step size and improve the stability of the model.

[0013] To address the drawback of traditional location methods, which are limited to single velocity assumptions, this paper combines a deep feedforward neural network model with a multi-head self-attention mechanism and a residual module for microseismic event location. This method first collects spatiotemporal information from multiple monitoring points (including the 3D coordinates of each monitoring point and the arrival time of each received seismic signal) as input data. The arrival time differences are calculated and reorganized into a feature matrix, which is then normalized to eliminate dimensional differences. The multi-head self-attention mechanism is then used to split the spatiotemporal information into subspaces, assigning self-attention weights to each subspace and capturing the correlation between different feature dimensions. Furthermore, a residual module is used to prevent gradient vanishing, extracting spatial time differences from the layered velocity model data, and train the deep feedforward neural network model to ensure its training stability. Finally, the model output layer outputs the predicted 3D coordinates of the earthquake source. Experimental results show that compared with the traditional FNN model, the Geiger location method, and the Newton iteration method, the proposed method reduces the average location error by 71.4%, 56.7%, and 66.4%, respectively. This method offers advantages such as high positioning accuracy and high computational efficiency, making it particularly suitable for real-time monitoring and location of microseismic events in complex geological environments with layered velocity models. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is the overall flow chart of the present invention; Figure 2 Schematic diagram of the residual module in the present invention; Figure 3 The kernel density distribution diagram of positioning error of the four methods; Among them, (a) is the positioning error of the embodiment of the present invention; (b) is the positioning error of the FNN model; (c) is the positioning error of the Geiger positioning method; (d) is the positioning error of the Newton iteration method. DETAILED DESCRIPTION

[0015] The present invention will be further described below.

[0016] like Figure 1 As shown, the present invention includes the following steps: S1. Select the spatiotemporal information of four monitoring points (including the three-dimensional coordinates of each monitoring point and the arrival time of each receiving the earthquake signal, i.e., (xi, yi, zi, ti)) as input sample data; at the same time, obtain the actual source position, i.e., (x, y, z), as verification sample data; the present invention only requires the spatiotemporal information of four monitoring points to achieve accurate positioning of the microseismic source. Compared with other existing methods, fewer monitoring points are used, which facilitates data collection and processing.

[0017] S2. Calculate the arrival time difference of earthquake signals received between monitoring points , replace the original monitoring arrival data to form a 1×16 one-dimensional array; the arrival time difference The specific expression is: Where, Indicates the monitoring time of the i-th monitoring point in the input data, i=1,…,4; Indicates the arrival time difference relative to the first monitoring point.

[0018] S3. Reshape the one-dimensional array in step S2 into a 4×4 two-dimensional feature matrix using the Reshape function. Specifically: Considering that the coordinate dimension and time dimension of the monitoring point are m and s respectively, the difference between the numerical scales is large. In order to avoid the problem of excessive dependence of the network model on a certain feature due to numerical imbalance, the mean is used. and standard deviation , standardize different features and balance the values ​​of input data, specifically: Where, is the reorganized two-dimensional feature matrix; For the The monitoring point corresponds to the The standard value of a feature.

[0019] S4. Flatten the normalized feature matrix into a 16-dimensional vector through the Flatten layer of the deep feedforward neural network , thereby adapting the receiving form of the hidden layer and using the hidden layer to convert the data into a 32-dimensional feature vector with higher discrimination. The specific formula is: Where, is the weight matrix with a dimension of 32×16; is the flattened 16-dimensional vector; is a 32-dimensional bias vector; is a 32-dimensional feature vector.

[0020] During this processing step, the coordinate and time information in the data are in an associated space. Combined with the ReLU activation function, the mapping relationship between data features can be captured. Coordinate and time features with small differences can be mapped to similar vectors, and vice versa, they can be mapped to vectors with more significant differences, thereby improving the model's ability to recognize different spatial and temporal difference features in subsequent work.

[0021] S5. Convert the 32-dimensional feature vector with higher discrimination in step S4 into a 4×8 feature matrix through the Reshape function and pass it along the two branches: The first branch: extract global dependencies through the multi-head self-attention mechanism and process them through the normalization layer. Specifically, the multi-head self-attention mechanism adjusts the dimension of the input matrix on top of the self-attention mechanism, thereby changing the structural dimension of the query matrix Q, key matrix K and value matrix V. Specifically, self-attention will split the input feature matrix into multiple subspaces with the same number of monitoring points. Each subspace focuses on different types of associated data, corresponding to the three directional coordinates and monitored time difference between each monitoring point; then, through 、 、 Perform weighted summation in each subspace, assign attention weights, and obtain the splicing results of each subspace, so as to extract the global dependency relationship of features from multiple angles. The specific expression is: Where, is the subspace obtained by splitting the input feature matrix, ; is the output of the multi-head self-attention layer, with a dimension of 4×8; 、 and They are 、 and The corresponding weight.

[0022] The second branch: acts as a skip connection and retains the original feature matrix.

[0023] S6. Combine the outputs of the two branches in the concatenation layer of the deep feedforward neural network to obtain the concatenation result. , and then the specific expression of deep fusion of features using the hidden layer of 256 neurons is: At the same time, in order to prevent the model from overfitting due to an excessive number of neurons, a Dropout module is added in the process to shield the output of some neurons. The specific expression is: S7, the fusion result of step S6 is processed by two layers of residual modules in sequence. Figure 2 As shown in the figure, each layer of the residual module contains two fully connected layers (FC) and two batch normalization layers (BN). The spatial time difference regularity is extracted by stacking two layers of residual modules. The specific calculation process of each layer of the residual module is: Where, and represent the input and output of the residual block respectively.

[0024] S8. After training the deep feedforward neural network model according to the spatial time difference law extracted in step S7, its output layer outputs the predicted three-dimensional coordinates of the earthquake source through 3 neurons, and calculates the loss based on the verification sample data in step S1 to optimize the deep feedforward neural network model. The specific optimization process is as follows: the verification sample data is divided into training set and test set in a ratio of 8:2, and the network model uses the trainingOptions function built into MATLAB to configure hyperparameters; through multiple rounds of hyperparameter adjustment, a reasonable mini-batch size (MiniBatchSize) and maximum training cycle (MaxEpochs) are set to optimize the memory usage of the model; at the same time, the initial learning rate is set to 0.0005, and the learning rate is halved every 5 data traversals to shorten the training step size and improve the stability of the model.

[0025] S9. When the latest earthquake signal received by the four monitoring points in step S1 arrives, it is input into the model optimized in step S8, and the model outputs the three-dimensional coordinates of the earthquake source corresponding to the current earthquake signal.

[0026] In order to verify the accuracy of microseismic source positioning of the present invention, a complex geological environment of a layered wave velocity model is first set up, and then four monitoring points are arranged and their position information is obtained. Then, seismic waves are generated at known locations as microseismic sources, and the four monitoring points receive and determine the arrival time of their respective seismic signals. Then, the present invention is combined with the existing FNN model, Geiger positioning method and Newton iteration method, all using the above four monitoring points and the corresponding seismic signal arrival time to calculate the three-dimensional coordinates of the microseismic source. The positioning results are shown in Figure 2. Figure 3As shown; finally, the positioning results of each method are compared with the actual three-dimensional coordinates of the microseismic source. Experiments show that compared with the traditional FNN model, Geiger positioning method and Newton iteration method, the average positioning error of the present invention is reduced by 71.4%, 56.7% and 66.4% respectively. It has the advantages of high positioning accuracy and fast calculation efficiency, and is particularly suitable for real-time monitoring and positioning of microseismic events in complex geological environments with layered wave velocity models.

[0027] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A microseismic intelligent positioning method based on deep feedforward neural network, characterized in that: The following steps are involved: S1. Select the spatiotemporal information of multiple monitoring points as input sample data; and simultaneously obtain the actual earthquake source location as verification sample data; S2. Calculate the arrival time difference of earthquake signals received between monitoring points , replace the original monitoring arrival data and form a one-dimensional array; S3, reorganize the one-dimensional array of step S2 into a two-dimensional feature matrix, and then use the mean and standard deviation , standardize different features and balance the values ​​of input data; S4. Flatten the normalized feature matrix through the Flatten layer of the deep feedforward neural network to adapt it to the receiving form of the hidden layer, and use the hidden layer to convert the data into a feature vector with higher discrimination; S5: Convert the more discriminative feature vectors from step S4 into feature matrices and pass them along two branches: the first branch extracts global dependencies through a multi-head self-attention mechanism and processes them through a normalization layer; the second branch acts as a skip connection and retains the original feature matrix. S6. The outputs of the two branches are merged in the concatenation layer of the deep feedforward neural network and deeply fused through the hidden layer. At the same time, a Dropout module is added in this process to mask the outputs of some neurons. S7, passing the fusion result of step S6 to the residual module, and extracting the spatial time difference regularity through the stacked two layers of residual modules; S8. After training the deep feedforward neural network model according to the spatial time difference rule extracted in step S7, its output layer outputs the predicted three-dimensional coordinates of the earthquake source, and calculates the loss based on the verification sample data in step S1 to optimize the deep feedforward neural network model; S9. When the latest earthquake signals received by multiple monitoring points in step S1 are obtained, they are input into the model optimized in step S8, and the model outputs the three-dimensional coordinates of the earthquake source corresponding to the current earthquake signal.

2. The microseismic intelligent positioning method based on deep feedforward neural network according to claim 1 is characterized in that: The number of monitoring points selected in step S1 is four.

3. The microseismic intelligent positioning method based on deep feedforward neural network according to claim 2 is characterized in that: The two-dimensional feature matrix in step S3 is specifically: Using the mean and standard deviation , standardize different features, specifically: Where, is the reorganized two-dimensional feature matrix; For the The monitoring point corresponds to the The standard value of a feature.

4. The microseismic intelligent positioning method based on deep feedforward neural network according to claim 1 is characterized in that: During the processing of step S4, the coordinate and time information in the data are in an associated space, and combined with the ReLU activation function, the mapping relationship between data features can be captured; coordinate and time features with smaller differences can be mapped to similar vectors, and vice versa, they are mapped to vectors with more significant differences, thereby improving the model's ability to recognize different spatial and temporal difference features in subsequent work.

5. The microseismic intelligent positioning method based on a deep feedforward neural network according to claim 1 is characterized in that: The multi-head self-attention mechanism in step S5 is based on the self-attention mechanism, and adjusts the dimension of the input matrix, thereby changing the structural dimension of the query matrix Q, the key matrix K and the value matrix V. Specifically, the self-attention will split the input feature matrix into multiple subspaces with the same number of monitoring points. Each subspace focuses on different types of associated data, corresponding to the three directional coordinates and the monitored time difference between each monitoring point; then, through 、 、 Perform weighted summation in each subspace, assign attention weights, and obtain the splicing results of each subspace, so as to extract the global dependency relationship of features from multiple angles. The specific expression is: Where, The subspace obtained by splitting the input feature matrix; is the output of the multi-head self-attention layer; 、 and They are 、 and The corresponding weight.

6. The microseismic intelligent positioning method based on deep feedforward neural network according to claim 1 is characterized in that: In step S7, the residual module has two layers, and the specific calculation process of each residual module is as follows: Where, and represent the input and output of the residual block respectively.

7. The microseismic intelligent positioning method based on deep feedforward neural network according to claim 1 is characterized in that: In step S8, the loss is calculated to optimize the deep feedforward neural network model. The specific optimization process is as follows: the verification sample data is divided into a training set and a test set in a ratio of 8:2, and the network model uses the trainingOptions function built into MATLAB to configure hyperparameters; through multiple rounds of hyperparameter adjustment, a reasonable mini-batch size and maximum training cycle are set to optimize the memory usage of the model; at the same time, the initial learning rate is set to 0.0005, and the learning rate is halved every 5 data traversals to shorten the training step size and improve the stability of the model.

Images