Microseismic p-wave and s-wave separation method based on cnn-lstm and multi-head attention mechanism

Abstract

The present application belongs to the field of microseismic monitoring, and relates to a microseismic P wave and S wave separation method based on CNN-LSTM and a multi-head attention mechanism. The technology comprises: constructing a data set, using microseismic data collected in a hot dry rock hydraulic fracturing field, and manually labeling the data; the data set includes microseismic data of different time periods and different background noise intensities; inputting the data set into a neural network composed of a convolution layer, an excitation layer, a pooling layer, a random hidden neuron (Dropout) layer, a long short-term memory (LSTM) layer, a multi-head attention mechanism layer and a full connection layer for training; after the training is completed, the model is saved; inputting the microseismic data to be separated into the trained model for separation, and finally obtaining the separation result of P wave and S wave.

Description

Technical Field

[0001] This invention belongs to the field of microseismic monitoring, and specifically relates to a microseismic P-wave and S-wave separation method based on CNN-LSTM and multi-head attention mechanism. Background Technology

[0002] Currently, the key technology for developing hot dry rock resources is hydraulic fracturing, which involves injecting high-pressure fluid into the reservoir to create complex artificial fractures, thereby extracting heat from the hot dry rock. During hydraulic fracturing operations, problems often arise such as difficulty in locating the fracture, difficulty in observing the magnitude of events, and unclear spatial orientation of the fractures. To address these issues, monitoring and analyzing the entire hydraulic fracturing process using the numerous observable microseismic events generated during fracturing becomes crucial, serving as a prerequisite for efficient and safe development of hot dry rock resources.

[0003] Microseismic monitoring technology mainly includes effective event picking, effective event signal enhancement, and microseismic source location. In actual monitoring, situations with low signal-to-noise ratios are often encountered. Because microseismic signals are weak, they are easily drowned out by background noise, making accurate identification impossible and affecting subsequent analysis. For example, when locating microseismic sources, their high sensitivity to the arrival times of P-waves and S-waves means that an error of 0.01 seconds can translate into a significant location error, hindering guidance for hydraulic fracturing and the optimization of engineering parameters.

[0004] Typically, microseismic monitoring is a long-term and continuous process, and the data acquired by detectors is often in the terabytes (TB) range. In such cases, manual signal acquisition or conventional methods cannot guarantee the real-time monitoring needs of microseismic events.

[0005] In conclusion, proposing a method that can overcome noise interference and accurately and efficiently separate P-waves and S-waves under on-site construction conditions is of great significance. Summary of the Invention

[0006] To address the aforementioned technical problems in existing technologies, this invention provides a microseismic P-wave / S-wave separation method based on CNN-LSTM and multi-head attention mechanism.

[0007] This invention is implemented as follows:

[0008] A microseismic P-wave / S-wave separation method based on CNN-LSTM and multi-head attention mechanism includes the following:

[0009] S1: 10,000 microseismic data channels were obtained by using effective microseismic data collected on-site from hot dry rock hydraulic fracturing. Each microseismic data channel contains P-wave and S-wave signals.

[0010] S2: Preprocess each microseismic data point by cutting the data length to 5000 sampling points and manually labeling the data to form the training dataset.

[0011] S3: Build a neural network model consisting of convolutional layers, activation layers, pooling layers, random hidden neuron layers, long short-term memory layers, multi-head attention mechanism layers, and fully connected layers, and start training using the training dataset in step S2;

[0012] S4: After training is complete, save the neural network model, input the data to be separated into the model for separation, and output the probability value of whether each sampling point is a P-wave or an S-wave, and finally obtain the separation result of P-wave and S-wave.

[0013] Furthermore, step S2 specifically includes:

[0014] The collected data is truncated to a uniform length. P-wave or S-wave positions in the data are marked as 1, and other positions are marked as 0.

[0015] Furthermore, S3 specifically includes:

[0016] The neural network model, from input to output, consists of convolutional layers, activation layers, pooling layers, dropout layers, LSTM layers, multi-head attention mechanisms, and fully connected layers. The convolutional kernel parameters in the convolutional layers are set to 128×16×1, 64×32×1, 32×32×1, and 16×16×1, respectively. The activation functions used in the activation layers are Sigmoid, Tanh, and ReLU. Max pooling is used in the pooling layers. The dropout probability in the dropout layers is set to 0.3. The input size and hidden size of the LSTM layers are both set to 16. The total size of the multi-head attention mechanism layer is set to 32, and the number of heads in the multi-head attention mechanism is set to 4. The fully connected layers have 32 input features and 1 output feature. The mean squared loss function is used.

[0017] loss(x i ,y i )=(x i -y i ) 2

[0018] In the formula x i y represents the predicted value output by the model. i The label represents the actual data.

[0019] Compared with the prior art, the beneficial effects of this invention are as follows:

[0020] This invention utilizes a Convolutional Neural Network (CNN) to extract and learn high-latitude features from microseismic data, effectively overcoming the impact of background noise on the accuracy of P-wave / S-wave separation. Combining a Long Short-Term Memory (LSTM) model with the CNN allows the model to consider the correlation between preceding and subsequent data, eliminating the adverse effects of irrelevant data on the separation results. A multi-head attention mechanism is introduced on top of the CNN-LSTM model. Microseismic data acquired in real-time at hot dry rock hydraulic fracturing sites typically contains significant background noise, which substantially impacts the accuracy of P-wave / S-wave separation. Therefore, the multi-head attention mechanism reduces the model's requirements on the signal-to-noise ratio of the original data, minimizing the need for preliminary data processing. The multi-head attention mechanism increases the weight of important information and decreases the weight of unimportant information in the microseismic data, enabling the model to learn which parts of the microseismic data contain crucial information for P-wave / S-wave separation. This allows the model to improve the accuracy of P-wave / S-wave separation even under background noise interference.

[0021] Compared with traditional methods, this invention eliminates the need for manual parameter changes or threshold settings after network training, facilitating later deployment and real-time field use. Validation was performed using monitoring data from hydraulic fracturing in hot dry rock in Qinghai Province, demonstrating that it can rapidly separate P-waves and S-waves while maintaining high separation accuracy.

[0022] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. Attached Figure Description

[0023] Figure 1 A flowchart of the method of the present invention is shown;

[0024] Figure 2 The label styles in the dataset are shown in an embodiment of the present invention;

[0025] Figure 3 The diagram shows the model structure of CNN-LSTM and multi-head attention mechanism in this invention;

[0026] Figure 4 The actual separation results of P-wave and S-wave using the method of the present invention are shown; Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0028] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0029] Figure 1 The flowchart shown illustrates the method of this invention. It includes constructing a dataset, building and training a CNN-LSTM model, saving the model, and separating microseismic data using the model. Specifically:

[0030] A microseismic P-wave / S-wave separation method based on CNN-LSTM and multi-head attention mechanism includes the following:

[0031] S1: 10,000 microseismic data channels were obtained by using effective microseismic data collected on-site from hot dry rock hydraulic fracturing. Each microseismic data channel contains P-wave and S-wave signals.

[0032] S2: Preprocess each microseismic data point by cutting the data length to 5000 sampling points and manually labeling the data to form the training dataset.

[0033] S3: Build a neural network model consisting of convolutional layers, activation layers, pooling layers, random hidden neuron layers, long short-term memory layers, multi-head attention mechanism layers, and fully connected layers, and start training using the training dataset in step S2;

[0034] S4: After training is complete, save the neural network model, input the data to be separated into the model for separation, and output the probability value of whether each sampling point is a P-wave or an S-wave, and finally obtain the separation result of P-wave and S-wave.

[0035] Step S2 specifically includes: the collected data is truncated to a uniform length, and the P-wave or S-wave position in the data is marked as 1, and the other positions are marked as 0.

[0036] S3 specifically includes the following neural network model from input to output: convolutional layers, activation layers, pooling layers, dropout layers, LSTM layers, multi-head attention mechanisms, and fully connected layers. The convolutional kernel parameters in the convolutional layers are set to 128×16×1, 64×32×1, 32×32×1, and 16×16×1, respectively. The activation functions used in the activation layers are Sigmoid, Tanh, and ReLU. Max pooling is used in the pooling layers. The dropout probability in the dropout layers is set to 0.3. The input size and hidden size of the LSTM layers are both set to 16. The total size of the multi-head attention mechanism layer is set to 32, and the number of heads in the multi-head attention mechanism is set to 4. The input features of the fully connected layers are set to 32, and the output features are set to 1. The mean squared loss function is used.

[0037] loss(x i ,y i )=(x i -y i ) 2

[0038] In the formula x i y represents the predicted value output by the model. i The label represents the actual data.

[0039] Figure 2 As shown, this invention's data set has a centralized label style. P-wave and S-wave positions are labeled as 1, and all other positions are labeled as 0.

[0040] Figure 3 The diagram illustrates the specific structure of the CNN-LSTM and multi-head attention mechanism model in this invention. The input data dimension is 5000×1, and the input label dimension is also 5000×1. Layers 1, 2, 3, and 4 are convolutional layers with kernel sizes of 128×16×1, 64×32×1, 32×32×1, and 16×16×1, respectively. Each convolutional layer is followed by an activation function: the first layer uses a sigmoid function to place the input in the 0-1 range to prevent ReLU activation; layers 2, 3, and 4 use ReLU functions to address the vanishing gradient problem; the fifth layer is a pooling layer, using max pooling; and the sixth layer is a dropout layer with a dropout probability set to 0. 3; Layer 7 is an LSTM layer with an input size of 16 and a hidden size of 16; Layer 8 is a multi-head attention layer with a total size of 32 and 4 heads; Layer 8 is followed by a Tanh function as an activation function to change the output to be centered at 0, so as to facilitate the final probability output of the model; Layer 9 is a fully connected layer with an input feature of 32 and an output feature of 1; Layer 9 is followed by a Tanh function as an activation function for the same reason as Layer 8.

[0041] Figure 4 The actual separation results of the P-wave and S-wave described in this invention are shown. The results are validated using data collected from actual hydraulic fracturing in hot dry rock in Qinghai. When the probability value of either the P-wave or S-wave is continuously greater than 0.5, it indicates successful separation of the P-wave or S-wave. Probability values ​​continuously greater than 0.5 are set to 1, and all others are set to 0 for observation. The experimental results show that the model performs well in separating the P-wave and S-wave, and has high accuracy in picking up the initial arrival position.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A microseismic P-wave / S-wave separation method based on CNN-LSTM and multi-head attention mechanism, characterized in that, Including the following: S1: 10,000 microseismic data channels were obtained by using effective microseismic data collected on-site from hot dry rock hydraulic fracturing. Each microseismic data channel contains P-wave and S-wave signals. S2: Preprocess each microseismic data point by cutting the data length to 5000 sampling points and manually labeling the data to form the training dataset. S3: Build a neural network model consisting of convolutional layers, activation layers, pooling layers, random hidden neuron layers, long short-term memory layers, multi-head attention mechanism layers, and fully connected layers, and start training using the training dataset in step S2; S4: After training is complete, save the neural network model, input the data to be separated into the neural network model for separation, and output the probability value of whether each sampling point is a P-wave or an S-wave, and finally obtain the separation result of P-wave and S-wave. S3 specifically includes: The neural network model, from input to output, consists of convolutional layers, activation layers, pooling layers, dropout layers, LSTM layers, multi-head attention mechanisms, and fully connected layers. The convolutional kernel parameters in the convolutional layers are set to 128×16×1, 64×32×1, 32×32×1, and 16×16×1. The activation functions used in the activation layers are Sigmoid, Tanh, and ReLU. Max pooling is used in the pooling layers. The dropout probability in the dropout layers is set to 0.

3. The input size and hidden size of the LSTM layers are both set to 16. The total size of the multi-head attention mechanism layer is set to 32, and the number of heads in the multi-head attention mechanism is set to 4. The fully connected layers have 32 input features and 1 output feature. The mean squared loss function is used. ， In the formula This represents the predicted value output by the model. The label represents the actual data.

2. The microseismic P-wave / S-wave separation method based on CNN-LSTM and multi-head attention mechanism according to claim 1, characterized in that, Step S2 specifically includes: The collected data is truncated to a uniform length. P-wave or S-wave positions in the data are marked as 1, and other positions are marked as 0.

Images