A microseismic P-wave first arrival picking method based on UNet3+

Abstract

The present invention discloses a microseismic P-wave first-arrival picking method based on UNet3+. First, a microseismic waveform database is established; then, the processed microseismic waveform is normalized; finally, a P-wave arrival time picking model is constructed with the help of the U‑Net3+ deep learning network, and the P-wave arrival time picking effect is tested. The present invention uses deep learning to guide the first-arrival picking of microseismic signals, realizes efficient first-arrival picking of massive microseismic signals; realizes faster and more accurate first-arrival picking of microseismic signals, further enhances the signal-to-noise discernibility of microseismic signals, and improves the accuracy of first-arrival picking of microseismic signals. The present invention can effectively improve the efficiency and accuracy of P-wave selection in microseismic data analysis, has great potential in the practical application of microseismic monitoring and prediction, has strong practical significance for promoting the application of deep learning in automatic first-arrival picking of P-waves, and can provide an important theoretical basis for serving the country's energy strategic security.

Description

Technical Field

[0001] The present invention relates to the field of microseismic monitoring technology, and in particular to a microseismic P wave first arrival picking method based on UNet3+. Background Art

[0002] Microseismic monitoring is a rapidly developing reservoir engineering management tool that can be used for reservoir fracturing monitoring and reservoir dynamic monitoring during oil and gas reservoir exploration and production, and has been widely used in the field of hydraulic fracturing monitoring. The advantage of microseismic monitoring technology compared to other monitoring technologies is that it can understand the direction of fractures and evaluate the effect of fracturing by locating microseismic events in time and space, and assist reservoir engineers in continuously adjusting and optimizing fracturing design plans. Therefore, accurate positioning of microseismic events is particularly important, and is also the focus and difficulty of current microseismic monitoring technology.

[0003] At present, the factors that affect the accuracy of microseismic positioning include arrival time picking, positioning algorithm and velocity model selection. Among them, arrival time picking is the basis for analyzing and processing microseismic signals, and its accuracy directly affects the location of the earthquake source, crack prediction, analysis of the earthquake source mechanism, etc. However, in current engineering applications, the traditional picking methods generally used (such as STA / LTA, AIC criterion, template matching) often require manual setting of thresholds, which are greatly affected by the thresholds, and the picking efficiency is low, which makes it difficult to meet the needs of first arrival picking of microseismic signals in the big data era. Summary of the invention

[0004] The purpose of the present invention is to improve the accuracy and efficiency of picking the first arrival of microseismic P waves, and to provide a microseismic P wave first arrival picking method based on UNet3+.

[0005] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0006] The present invention comprises the following steps:

[0007] Step 1: Use the three-component microseismic data with manually annotated P-wave arrival times in the STEAD dataset with Magnitude ≤ 4, Straight-light Distance ≤ 25 km, and Depth / Distance ≥ 0.3 as the original data for preprocessing;

[0008] Step 2: Pick the original data set for the first time, process the labels and prediction results, and randomly divide the samples in the data set into training set and test set according to the preset ratio.

[0009] Step 3: Add Residual blocks and CBAM blocks to the U-Net3+ network to build a phase feature extraction and fusion-enhanced arrival time picking network to train the microseismic signal data set; the network mainly consists of two parts: encoder and decoder. The encoder mainly extracts features from the input microseismic data, enhances the weight assigned to the first arrival of the phase in the sequence, and reduces the interference of other information on the first arrival picking; the decoder restores the extracted feature sequence to a sequence with the same length as the input data, and finally identifies each time step.

[0010] The fully convolutional residual module of the present invention is different from the common residual module. The present invention replaces the pooling layer of the original residual network with a convolution layer with a step size of 2 to minimize the loss of high-frequency feature information caused by using the pooling layer in the downsampling process;

[0011] Step 4: The convolutional block attention module (CBAM) shows remarkable effectiveness in processing sequence data of different lengths, making it very suitable for microseismic P-wave picking. It is known that a complete microseismic event data is composed of noise segments, P-wave segments, S-wave segments and noise segments in sequence in time series; and the signal-to-noise ratio of the signal on different components is different but the P-wave first arrival is the same. In view of these two characteristics, the CBAM attention mechanism is added to the network, hoping that the network will pay more attention to the components with higher signal-to-noise ratio and pay more attention to the segments near the P-wave first arrival. CBAM includes channel attention mechanism and spatial attention mechanism. First, the channel attention is multiplied by the original feature, and then the result is multiplied by the spatial attention to obtain the re-assigned feature.

[0012] Step 5: Predict the test set under different signal-to-noise ratios, and obtain the recognition accuracy and overall performance of the network model based on the evaluation indicators. Use Softmax to predict the "binary classification" probability of each sampling point of the signal to be tested belonging to the first arrival category and the non-first arrival category. Finally, the middle time step of the sequence with a probability sequence of 1 is used as the first arrival time of the P wave. If the error between the predicted first arrival time and the actual first arrival time is within 0.25s, the prediction is considered correct, and the accuracy rate (H) and the missed rate (R) are used to evaluate the algorithm;

[0013] Step 6: Since positive and negative samples are unbalanced, the loss function of the present invention adopts Focal loss, and the probability p of the predicted sample being positive is t The relationship with L is as follows:

[0014]

[0015] Where i represents the number of sampling points, γ is an adjustable parameter, and α is a vector representing the weight of each category.

[0016] Preferably, the preprocessing is to perform Min-Max normalization on the data to speed up the network convergence;

[0017] Preferably, the channel attention mechanism and the spatial attention mechanism described in step 4 include the formula:

[0018]

[0019] Where σ represents the sigmoid activation function, W0 and W1 represent two convolution operations respectively. and Represent average pooling and maximum pooling respectively;

[0020]

[0021] where f 7×7 Represents a convolution operation with a convolution kernel size of 7×7

[0022] Preferably, the Softmax formula described in step 5 is as follows: Where Z i is the output value of the i-th node, k is the number of nodes;

[0023] The formula for accuracy (H) is as follows: H = N / M, where N represents the number of microearthquake first arrivals correctly detected by the model, and M represents the total number of events;

[0024] The formula for missed pick-up rate (R) is as follows: R = K / M, where K represents the number of microearthquake first arrivals that the model did not detect;

[0025] Preferably, after step 5, an ablation experiment is performed and the results are analyzed and evaluated.

[0026] The beneficial effects of the present invention are:

[0027] The present invention is a microseismic P wave first arrival picking method based on UNet3+. Compared with the prior art, the present invention uses deep learning to guide the first arrival picking of microseismic signals, and realizes efficient first arrival picking of massive microseismic signals. In order to reduce the information loss in the process of extracting microseismic waveform features, the U-Net3+ decoder is enhanced, which is characterized by adopting a residual structure to better propagate information, replacing the pooling layer with a convolutional layer to retain the high-frequency microseismic waveform details, and adding a convolutional block attention module (CBAM) to the decoder and encoder in a deeper layer, so as to hope that the network pays more attention to the key information when the microseismic wave first arrives, and realize faster and more accurate microseismic first arrival picking, further enhance the signal-noise discernibility of microseismic signals, and improve the accuracy of microseismic first arrival picking. The present invention can effectively improve the efficiency and accuracy of P wave selection in microseismic data analysis, has great potential in the practical application of microseismic monitoring and prediction, has strong practical significance for promoting the application of deep learning in automatic first arrival picking of P waves, and can provide an important theoretical basis for serving the national energy strategic security. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the method for automatically picking up the P wave arrival time based on U-Net3+ in the present invention;

[0029] Figure 2 It is a schematic diagram of the overall structure of the network in the present invention;

[0030] Figure 3 This is a schematic diagram of the structure of the Residual module used in the present invention. The numbers in brackets represent the convolution kernel size, step size, and padding, respectively.

[0031] Figure 4 This is a schematic diagram of the CBAM module structure used in the present invention;

[0032] Figure 5 This is a graph showing the results of the present invention in picking up microseismic signals with a signal-to-noise ratio (SNR) of less than 10 dB;

[0033] Figure 6 This is a diagram showing the results of the present invention in picking up microseismic signals with a signal-to-noise ratio (SNR) of less than 5 dB;

[0034] Figure 7 It is a bar chart comparing the error of the model in the present invention for the P-wave arrival time picking results of microseismic signals with different signal-to-noise ratios;

[0035] Figure 8 It is a test diagram of the anti-noise performance of the model in the present invention and other pickup networks. DETAILED DESCRIPTION

[0036] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0037] In view of the deficiency that the existing microseismic first-arrival picking technology is difficult to efficiently and accurately pick up massive microseismic P-wave first arrivals, the present invention combines deep learning technology and applies the classification method to velocity model selection, and provides a microseismic P-wave first-arrival picking method based on U-Net3+, which improves the first-arrival picking accuracy of microseismic signals.

[0038] First, a microseismic waveform database is established; then, the processed microseismic waveforms are normalized; finally, a P-wave arrival picking model is constructed with the help of the U-Net3+ deep learning network, and the P-wave arrival picking effect is tested. The P-wave arrival picking method based on U-Net3+ proposed in the present invention can be well matched with the manual picking thinking, greatly improving the picking efficiency and accuracy of the P-wave first arrival of hydraulic fracturing microseismic signals, and laying the foundation for automatic, high-precision positioning calculation of microseisms and focal mechanism inversion.

[0039] The invention is further described below with reference to the accompanying drawings and specific implementation examples:

[0040] Step 1: Use the three-component microseismic data of manually annotated P-wave arrival time in the STEAD dataset with Magnitude≤4, Straight-light Distance≤25km, and Depth / Distance≥0.3 as the original data and take 22,909 data from the original data as experimental data for preprocessing. The preprocessing is to perform Min-Max normalization on the data to speed up the network convergence.

[0041] Step 2: Expand the annotation range of the P wave arrival to 25 time steps centered on the manually annotated moment for labeling and prediction results, and randomly divide the samples in the data set into a training set and a test set. Since the P wave arrival is only a unit time step in the data sequence, it accounts for a very small proportion compared to the sequence containing 6000 moments, which is not conducive to feature learning of the network, so it is necessary to expand the annotation range of the P wave arrival. After the samples in the data set are randomly divided, the training set contains 13024 data for the training of the deep neural network model, the validation set contains 3200 data, and the test set contains 3200 data with a signal-to-noise ratio below 20dB, 2714 data with a signal-to-noise ratio below 10dB, and 771 data with a signal-to-noise ratio below 5dB, which are used to test and evaluate the recognition effect of the model. When the network completes the recognition of the input data, it outputs an arrival probability sequence of the same length as the data, and takes the middle time step of the sequence of 20 consecutive 1s in the sequence as the P wave arrival.

[0042] Step 3: Add Residual blocks and CBAM blocks to the U-Net3+ network to build a time-picking network with stronger phase feature extraction and fusion capabilities to train the microseismic signal data set. Construct a time-picking model based on the U-Net3+ network, which consists of an encoder and a decoder. A residual module is added to the encoder and decoder to improve the feature extraction capability; the encoder mainly extracts features from the input microseismic data, enhances the weight assigned to the first arrival of the phase in the sequence, and reduces the interference of other information on the first arrival picking; the decoder restores the extracted feature sequence to a sequence with the same length as the input data, and finally identifies each time step.

[0043] Step 4: In order to improve the recognition accuracy of the model and speed up the model training, the present invention only adds the CBAM mechanism after the fourth and fifth residual convolution and feature fusion stages. The CBAM module includes a channel attention mechanism and a spatial attention mechanism, which can make the network pay more attention to important features and improve the network's feature extraction ability. First, calculate the multiplication of the channel attention and the original feature, and then multiply the result by the spatial attention to obtain the re-assigned feature. The specific formula is as follows:

[0044]

[0045] Where σ represents the sigmoid activation function, W0 and W1 represent two convolution operations respectively. and Represent average pooling and maximum pooling respectively;

[0046]

[0047] where f 7×7 Represents a convolution operation with a convolution kernel size of 7×7

[0048] Step 5: Explore the effectiveness of each module and mechanism in the constructed network, conduct a series of ablation experiments, and analyze and evaluate the experimental results. For the main modules of the present invention, the present invention constructs multiple networks to verify the effectiveness of each module, and finally proves that the final module of the present invention has stronger on-time pickup performance.

[0049] Step 6: Predict the test set under different signal-to-noise ratios, and obtain the recognition accuracy and overall performance of the network model based on the evaluation indicators. Use Softmax to predict the "binary classification" probability of each sampling point of the signal to be tested belonging to the first arrival category and the non-first arrival category. Finally, the middle time step of the sequence with a probability sequence of 1 is used as the first arrival time of the P wave. If the error between the predicted first arrival time and the actual first arrival time is within 0.25s, the prediction is considered correct, and the accuracy rate (H) and the missed rate (R) are used to evaluate the algorithm; the Softmax formula is as follows: Where Zi is the output value of the i-th node, k is the number of nodes;

[0050] The formula for accuracy (H) is as follows: H = N / M, where N represents the number of microearthquake first arrivals correctly detected by the model, and M represents the total number of events;

[0051] The formula for missed pick-up rate (R) is as follows: R = K / M, where K represents the number of microearthquake first arrivals that the model did not detect;

[0052] Step 7: Due to the imbalance of positive and negative samples, the loss function of the present invention adopts Focal loss, and the probability p of the predicted sample being positive is t The relationship with L is as follows:

[0053]

[0054] Where i represents the number of sampling points, γ is an adjustable parameter, and α is a vector representing the weight of each category.

[0055] The UNet++ network uses the UNet3+ network technology in the paper "H.Huang, L.Lin, R.Tong, H.Hu, Q.Zhang, Y.Iwamoto, X.Han, Y.Chen, and J.Wu.2020."UNet 3+:A Full-Scale Connected UNet for Medical ImageSegmentation"; CBAM uses the technology in the paper "S.Woo, J.Park, J.Lee and I.Kweon,"CBAM:Convolutional Block Attention Module"".

[0056] like Figure 5-6 The figure shows the schematic diagram of the first arrival picking result of microseismic signals based on U-Net3+ in practical applications; Figure 5 This is the data with a signal-to-noise ratio below 10dB. Figure 6 This is data with a signal-to-noise ratio below 5dB.

[0057] Figure 7 The error bar comparison chart of the model in the present invention for the P wave arrival time picking results of microseismic signals with different signal-to-noise ratios shows that most errors are concentrated between 0-0.05s, accounting for 52.22% of data below 20db, 49.34% of data below 10db, and 47.33% of data below 5d. This shows that the model has a high model picking accuracy and can completely replace manual picking.

[0058] Figure 8Figure 1 is a test chart of the noise resistance performance of the model in the present invention and other picking networks. In order to evaluate the noise tolerance performance of the network of the present invention, Gaussian white noise was added to the selected microseismic signal. Subsequently, the present invention created 11 test sets with SNR ranging from -10dB to 10dB. The present invention used these 11 test sets to compare the performance of the network of the present invention with other P-wave picking networks (i.e., U-Net, U-Net2+, U-Net3+, and Phase Net). Figure 8 It shows that when the signal-to-noise ratio is between -10dB and 10dB, the network of the present invention achieves a much higher accuracy than other networks. When the signal-to-noise ratio is -4dB, the model of the present invention achieves an accuracy of 94.87%, exceeding the accuracy of U-Net, U-Net2+, U-Net3+ and PhaseNet, which are 88.82%, 89.03%, 89.59% and 85.1% respectively. In addition, the missed pick-up rate of the network of the present invention is almost 0%, while the missed pick-up rate of other networks gradually increases as the signal-to-noise ratio decreases. This result further confirms that the network of the present invention exhibits superior noise tolerance compared to other networks.

[0059] Table 1 is a comparison table of the accuracy and missed detection rate of the algorithm of the present invention and other algorithms; in data with different SNRs, compared with U-Net3+, the accuracy H of the method of the present invention is increased by 3.2%, 7.98% and 12.27% respectively; in terms of missed detection rate R, compared with U-Net3+, the method of the present invention is reduced by 100%, 97.38% and 96.96% respectively, indicating that the robustness of the method of the present invention is much better than that of the traditional method, proving that the method of the present invention has good perception ability and comprehensive picking performance for the characteristics of microseismic first arrival signals;

[0060] Table 1 Comparison of the accuracy and missed rate of the algorithm with other algorithms

[0061]

[0062]

[0063] Table 2 shows the ablation experiment results for each module of the network of the present invention. The proposed method is compared with seven variants: U-Net3+ without residual module, U-Net2+ without attention mechanism, and U-Net3+ with different attention mechanisms. The experimental results are listed in Table 2, where model 8 represents the network of the present invention. The results show that all components of the proposed method contribute to the overall performance; however, the Res+CBAM combination produces the best performance, with the lowest missed pick-up rate and the highest accuracy.

[0064] Table 2 Ablation experiment results for each module

[0065]

[0066] The technical solution of the present invention is not limited to the above-mentioned specific embodiments. All technical variations made according to the technical solution of the present invention fall within the protection scope of the present invention.

Claims

1. A microseismic P-wave first arrival picking method based on UNet3+, characterized in that: The following steps are involved: S1: Use the three-component microseismic data with manually annotated P-wave arrival times in the STEAD dataset with Magnitude≤4, Straight-light Distance≤25km, and Depth / Distance≥0.3 as the raw data for preprocessing; S2: Pick up the original data set for the first time, process the labels and prediction results, and randomly divide the samples in the data set into a training set and a test set according to a preset ratio; S3: Add the full convolution residual module and the convolution block attention module CBAM to the U-Net3+ network to build a time-picking network with stronger phase feature extraction and fusion capabilities to train the microseismic signal dataset; S4: In the fully convolutional Residual module, the pooling layer in the original residual network is replaced by a convolutional layer with a stride of 2 to minimize the loss of high-frequency feature information caused by the use of the pooling layer in the downsampling process; the convolutional block attention module CBAM includes a channel attention mechanism and a spatial attention mechanism. First, the channel attention is multiplied by the original feature, and then the result is multiplied by the spatial attention to obtain the re-assigned feature; S5: Predict the test set under different signal-to-noise ratios, and obtain the recognition accuracy and overall performance of the network model based on the evaluation indicators. Use Softmax to predict the "binary classification" probability of each sampling point of the signal to be tested belonging to the first arrival category and the non-first arrival category. Finally, the middle time step of the sequence with a probability sequence of 1 is used as the first arrival time of the P wave. If the error between the predicted first arrival time and the actual first arrival time is within 0.25s, the prediction is considered correct, and the accuracy rate H and the missed rate R are used to evaluate the algorithm. S6: Focal loss is used as the loss function, and the probability p of the predicted sample being positive is t The relationship with L is as follows: Where i represents the number of sampling points, γ is an adjustable parameter, and α is a vector representing the weight of each category.

2. The microseismic P wave first arrival picking method based on UNet3+ according to claim 1, characterized in that: The arrival time picking network in step S3 is composed of two parts: an encoder and a decoder. The encoder extracts features from the input microseismic data, enhances the weight assigned to the first arrival of the seismic phase in the sequence, and reduces the interference of other information on the first arrival picking; the decoder restores the extracted feature sequence to a sequence with the same length as the input data, and finally identifies each time step.

3. The microseismic P wave first arrival picking method based on UNet3+ according to claim 2 is characterized in that: The preprocessing in step S1 is to perform Min-Max normalization on the data to speed up the network convergence.

4. The microseismic P wave first arrival picking method based on UNet3+ according to claim 3 is characterized in that: The channel attention mechanism and the spatial attention mechanism described in step S4 include the formula: Where σ represents the sigmoid activation function, W0 and W1 represent two convolution operations respectively. and Represent average pooling and maximum pooling respectively; where f 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

5. The microseismic P wave first arrival picking method based on UNet3+ according to claim 4 is characterized in that: The Softmax formula in step S5 is as follows: Where Z i is the output value of the i-th node, k is the number of nodes; The formula for the accuracy rate H is as follows: H = N / M, where N represents the number of microseismic first arrivals correctly detected by the model, and M represents the total number of events; The formula for the missed rate R is as follows: R = K / M, where K represents the number of microearthquake first arrivals that were not detected by the model.

6. The microseismic P wave first arrival picking method based on UNet3+ according to claim 5 is characterized in that: After step S5 is completed, an ablation experiment is performed and the results are analyzed and evaluated.

Images