A method for predicting a mohorovicic discontinuity structure based on multi-task machine learning

Abstract

The present application relates to a kind of Moho structure prediction methods based on multi-task machine learning, comprising the following steps: S1, utilize the style migration technology based on convolutional neural network to learn noise feature from real data, real receive function noise style is migrated to simulation data, to generate noisy simulation data set;S2, utilize the multi-task convolutional neural network based on parameter sharing to complete prediction model training, obtain the prediction model that can predict the Moho structure below single station, wherein Moho structure includes depth, strike, dip angle;S3, using accelerated beckman and threshold iteration method to reconstruct in the actual prediction data of missing azimuth angle;S4, the Moho structure of single station is predicted.The method changes traditional thought and processing flow, combines deep neural network, realizes accurate prediction to earth deep Moho through clear, simple and concise process strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of deep crust structure research of the earth, and in particular to a Moho structure prediction method based on multi-task machine learning. BACKGROUND

[0002] The receiver function is a set of time series signals that can be effectively used to characterize the underground velocity interface, and is currently commonly used to study the structure of the underground sphere and to invert the S wave velocity by using its sensitivity to the velocity interface and seismic wave velocity. In addition, since the waveforms arranged by azimuth in each receiver function seismogram represent the waveforms penetrating from the deep interface and received from different anti-azimuth with the station as the center, the converted waves and multiple waves in the figure often change in amplitude and time due to the change in the position of the actual penetration point, and the waveforms and isochrones change with the anti-azimuth. The Ps converted wave and PpPS, PpSs+PsPs multiple waves in the radial receiver function are used to characterize the Moho interface signal together, and these signals contain the depth of the interface and the overall strike and dip angle information of the interface with the horizontal plane. However, the main characteristic information of the receiver function is mixed with the information reflecting the structure change, and the noise interference and anisotropy affect the waveform, making it very difficult to analyze the crustal structure below the station by simply interpreting the receiver function anti-azimuth characteristics. In this study, we hope to use the convolutional neural network to directly identify the anti-azimuth characteristics from the receiver function seismogram, construct the model, and predict the three-dimensional structure of the Moho interface below a single station.

[0003] The object of the present study is the middle-southern segment of the Tanlu fault zone and its adjacent area. The fault zone is the suture zone of the lower Yangzi plate and the North China plate, and also a deep-cutting lithospheric giant fault. Previous studies have used various methods to detect the crustal structure of the Tanlu fault zone and its adjacent area for a long time. The receiver function method reveals the overall structure of the regional Moho, and the seismic wave velocity and reflection seismic method also presents the intracrustal structure and velocity anomaly of the region. However, these previous techniques have obvious limitations. The use of stacking and interpolation may to some extent erase the detailed structural information presented by a single station, and the fixed-angle profile and excessive smoothing processing further cause the relief characteristics of the imaged crustal structure to become blurred or disappear. Therefore, our method only predicts the Moho structure below a single station, and obtains the Moho depth and overall interface strike and dip angle of each station. SUMMARY

[0004] The purpose of the present application is to solve the defects in the prior art and provide a Moho structure prediction method based on multi-task machine learning.

[0005] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0006] A Moho structure prediction method based on multi-task machine learning, comprising the following steps:

[0007] S1, learning noise features from real data using a convolutional neural network-based style transfer technology, migrating real receiver function noise style to simulated data, thereby generating a noisy simulated data set;

[0008] S2, using a multi-task convolutional neural network based on hard parameter sharing to complete the prediction model training, obtaining a prediction model that can predict the Moho structure under a single station, wherein the Moho structure includes depth, strike, and dip angle;

[0009] S3, using the accelerated Beckman and threshold iteration method to reconstruct the actual prediction data missing azimuth angle;

[0010] S4, predicting the Moho structure of a single station.

[0011] Further, the S1 specific operation is: first, a noise library is established, and each noise is randomly intercepted from the real receiver function to be predicted; for each real receiver function, a time is randomly selected, and a segment with the same length as the simulated data is intercepted as a noise sample;

[0012] Before adding noise to the 36 azimuth angle simulated data, 36 noise data are randomly extracted from the noise library to form a "style image", and a seismic map composed of simulated receiver functions arranged in reverse azimuth is "content image", a network used for data simulation adopts a VGG-19 network based on pre-trained and removed full connection layer, "content image" and "style image" are sent into the network for training, different loss functions are constructed according to the results output by the features (1), (2), (3), and then the reverse derivation is carried out, and the final simulated data with real noise is obtained after multiple iterations;

[0013] (1) content loss function:

[0014]

[0015] Respectively, the input image and the generated image after style transfer, here l represents the layer of the convolutional layer in the network, i represents the number of channels in the convolutional layer, and j represents the j th th position in the convolutional layer, F and P represent the feature responses output by the "content image" and "initial image" corresponding to the lth layer;

[0016] (2) style loss function:

[0017] First, the feature cross-correlation is defined using the Gram matrix:

[0018]

[0019] The style loss function is:

[0020]

[0021] Here, represents the input style image. M l is equal to the height x width of the output single feature map of the lth layer, N l represents the number of filters of the layer, A and G are the feature responses of the "style image" and the "initial image" corresponding to the feature cross-correlation of the lth layer, w l represents the weight of the lth layer.

[0022] (3) Total loss function:

[0023]

[0024] Here, alpha and beta are weight factors balancing the content loss and the style loss.

[0025] Further, based on the basic idea of multi-task learning of hard parameter sharing, the most intuitive and classic Share-Bottom in the multi-task machine learning algorithm is used to complete the model building, the depth is set as a classification task, the Moho structure and the dip angle are set as regression tasks, the loss function of the classification task is a cross-entropy loss function, the loss function of the regression task of the trend and the dip angle is a mean square error loss, and the total loss function is a linear addition of the three loss functions.

[0026] Finally, in actual data prediction, due to uneven distribution of earthquakes and poor quality of part of the anti-azimuth angle data, multiple azimuth angle continuous missing situations often occur in the real data. Therefore, before predicting the real data, an accelerated Beckman and threshold iteration combined method is used to reconstruct the data, the curvelet transform is used as a sparse base, the accelerated linear Beckman method is used, and the threshold iteration method is combined, and a new linear and exponential weighting factor is used to adjust the proportion of the accelerated Beckman method and the threshold iteration method. When continuous multiple channels are missing or the interval missing is less than 50%, the method can effectively complete the reconstruction.

[0027] The method changes the traditional idea and processing flow, combines the deep neural network, realizes accurate prediction of the deep Moho of the earth through a clear, simple and streamlined strategy, realizes multi-parameter prediction, and compared with the traditional method which can only calculate the Moho depth, the method can simultaneously predict the depth, the trend and the dip angle of the three Moho basic parameters. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 The figure is a flowchart of the present application.

[0029] Figure 2 A model architecture diagram for the simulation of the receiving function of the application;

[0030] Figure 3 A deep neural network model diagram for multi-task prediction of the application;

[0031] Figure 4 A prediction result diagram for the example of the application;

[0032] Figure 5 A comparison diagram of the receiving function reconstruction results of the BEB station. DETAILED DESCRIPTION

[0033] As shown in Figures 1 to 4 A Moho structure prediction method based on multi-task machine learning, comprising the following steps:

[0034] S1, using a style transfer technology based on a convolutional neural network to learn noise features from real data, and migrating the noise style of the real receiving function to the simulation data, thereby generating a noisy simulation data set;

[0035] S2, using a multi-task convolutional neural network based on hard parameter sharing to complete the prediction model training, and obtaining a prediction model that can predict the Moho structure under a single station, wherein the Moho structure includes depth, strike, and dip angle;

[0036] S3, using the accelerated Beckman and threshold iteration method to reconstruct the actual prediction data missing azimuth angle;

[0037] S4, predicting the Moho structure of a single station.

[0038] Further, the S1 specifically operates as follows: first, a noise library is established, and each noise is randomly intercepted from the real receiving function to be predicted. For each real receiving function, a time is randomly selected, and a fragment with the same length as the simulation data is intercepted as a noise sample;

[0039] Before adding noise to the 36 azimuth angle simulation data, 36 noise data are randomly extracted from the noise library to form a "style image". The seismic map composed of the simulation receiving function arranged in reverse azimuth is the "content image". The network used for data simulation adopts a VGG-19 network based on a pre-trained and removed full connection layer. The "content image" and "style image" are sent into the network for training. The results output by the features are constructed into different loss functions according to (1), (2), and (3). Then, the final simulation data with real noise is obtained through multiple iterations of backpropagation;

[0040] (1) Content loss function:

[0041]

[0042] are the input image and the generated image after style transfer, respectively, where l represents the layer number of the convolutional layer in the network, i represents the channel number in the convolutional layer, and j represents the jth position in the convolutional layer, F and P represent the feature responses corresponding to the outputs of the "content image" and the "initial image" of the lth layer, respectively. th

[0043] (2) Style loss function:

[0044] First, the Gram matrix is used to define the feature cross-correlation:

[0045]

[0046] The style loss function is:

[0047]

[0048] Here, a represents the input style image. M l is equal to the height x width of the output single feature map of the lth layer, N l represents the number of filters of the lth layer, A and G are the feature responses corresponding to the outputs of the "style image" and the "initial image" of the lth layer, respectively, and w l represents the weight of the lth layer.

[0049] (3) Total loss function:

[0050]

[0051] Here, a and b are weight factors for balancing the content loss and the style loss.

[0052] Further, based on the basic idea of multi-task learning of hard parameter sharing, the Share-Bottom, which is the most intuitive and classic algorithm in multi-task machine learning, is used to complete the model building. The depth is set as a classification task, and the Moho structure and dip angle are set as regression tasks. The loss function of the classification task is the cross-entropy loss function, the loss function of the regression tasks of the trend and dip angle is the mean square error loss, and the total loss function is the linear addition of the three loss functions.

[0053] ​Finally, in the actual data prediction, due to the uneven distribution of earthquakes and the poor quality of part of the anti-azimuth angle data, multiple azimuth angle continuous missing often occurs in the real data. Therefore, before predicting the real data, the method of combining accelerated Beckman and threshold iteration is used to reconstruct the data. The method uses curvelet transform as a sparse basis, uses accelerated linear Beckman method, and combines with threshold iteration method, uses new linear and exponential weighting factors to adjust the proportion of accelerated Beckman method and threshold iteration method. When the continuous multiple channel is missing or the interval missing is less than 50%, the method can effectively complete the reconstruction.

[0054] The application will be described in detail below in combination with the drawings and examples, as follows:

[0055] The China Earthquake Administration has built multiple 60s wideband fixed stations along the fault zone. These stations are monitored continuously for 24 hours a year, and are equipped with independent swing houses, with small environmental interference and sufficient data. We selected 10 wideband stations along the Tanlu fault zone as the data source, and intercepted the seismic event waveform data from 2013 to 2015. We selected teleseismic events with a magnitude greater than Ms5 and a hypocentral distance of 30-90°. Although there are teleseismic events distributed in most of the anti-azimuth angle interval, they are mainly concentrated in the circum-Pacific seismic belt. The data is preprocessed regularly: de-meaning, de-trending, filtering 0.02-2Hz, ray parameter dynamic correction, and calculating the radial receiver function by time domain deconvolution. Finally, the high-quality receiver functions are selected by visual inspection, and are stacked and arranged according to every 10° interval from Baz 0-360°. Since the original data corresponds to a time span of 4 years, in actual work, the average number of high-quality receiver functions used for stacking at each station is not less than 300, which ensures that the stacked data has obvious converted wave and multiple wave moveout.

[0056] Then the method of combining accelerated Beckman and threshold iteration is used to reconstruct the data. The number of iterations is set to 15, the scale of curvelet transform is 4, and the direction value is 16. Figure 5 In the figure, area a is the stacking result of the BEB station receiver function. The station has continuous missing between anti-azimuth angles 60-100° and discontinuous missing between anti-azimuth angles 330-360°. Figure 5 In the figure, area b is the result after reconstruction. The data missing between anti-azimuth angles 60-100° and 330-360° is well reconstructed, and the Ps converted wave is clearly arranged in the range of 4-5s, and the multiple wave can be clearly distinguished between 10s and 23s. Figure 2 All the denoised receiver functions are arranged, and the overall denoising effect is obvious from the figure, especially the noise suppression effect on both sides of the multiple wave, and the damage to the converted wave and multiple wave signals is small, and the multiple wave arrival time does not have obvious time shift.

[0057] Table 1 prediction results

[0058]

[0059] The specific operation of realizing the method is as follows:

[0060] 1. Using RAYSUM software, according to the average velocity, density and other parameters of P and S waves in the research area, the simulation data generation is completed.

[0061] 2. According to the style migration method, all the generated simulation data are simulated with noise.

[0062] 3. The simulation data are sent into the network to train the model by using the multi-task model framework proposed in the patent.

[0063] 4. The wideband seismic waveform is intercepted according to the event time, the standard pretreatment method (mean removal, trend removal, filtering and rotation) is used, the time domain deconvolution method is used to extract the receiver function, and the stacking is carried out according to every 10° azimuth interval.

[0064] 5. The data are reconstructed by using the accelerated Bergman and threshold iteration combined method.

[0065] 6. The actual data are predicted by using the generated model to obtain the deep crust structure of the station.

[0066] The method changes the traditional idea and processing flow, combines the deep neural network, realizes the accurate prediction of the deep Moho surface of the earth through the clear, simple and streamlined strategy, and realizes the prediction of multiple parameters. Compared with the traditional method which can only calculate the Moho surface depth, the present application can simultaneously predict the three basic parameters of Moho surface, i.e. depth, strike and dip angle.

[0067] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection claimed by the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the structure of the Moho surface based on multi-task machine learning, characterized in that, Includes the following steps: S1. Using convolutional neural network-based pattern transfer technology, noise features are learned from real data, and the noise patterns of the real receiving function are transferred to the simulated data, thereby generating a noisy simulated dataset. S2. A prediction model is trained using a multi-task convolutional neural network based on hard parameter sharing to obtain a prediction model that can predict the Moho structure below a single platform. The Moho structure includes depth, orientation, and downslope. Based on the basic idea of ​​multi-task learning with hard parameter sharing, the most intuitive and classic algorithm in multi-task machine learning, Share-Bottom, is used to build the model. Depth is set as the classification task, and the Moho structure and downslope are set as regression tasks. The loss function for the classification task is the cross-entropy loss function, and the loss function for the orientation and downslope regression tasks is the mean squared error loss. The total loss function is the linear sum of the three loss functions. S3. Reconstructing missing azimuth angles in actual prediction data using accelerated Beckman and threshold iteration methods. Due to uneven earthquake distribution and poor quality of some inverse azimuth angle data, multiple azimuth angles are often missing in the actual data. Therefore, before predicting the actual data, a combined accelerated Beckman and threshold iteration method will be used to reconstruct the data. This method uses curvelet transform as a sparse basis, uses the accelerated linear Beckman method, and combines it with the threshold iteration method. A novel linear and exponential weighting factor is used to adjust the proportion of the accelerated Beckman method and the threshold iteration method. S4. Predict the Moho surface structure of a single unit.

2. The method for predicting the Moho surface structure based on multi-task machine learning according to claim 1, characterized in that, The specific operation of S1 is as follows: First, a noise database is established. Each noise is randomly extracted from the real receiving function that needs to be predicted. For each real receiving function, a time is randomly selected, and a segment with the same length as the simulated data is extracted as a noise sample. Before adding noise to the 36 azimuth simulation data, 36 noise data are randomly selected from the noise library to form a "style image". The seismic map composed of simulated receiving functions arranged in reverse azimuth is the "content image". The network used for data simulation adopts a pre-trained VGG-19 network with the fully connected layers removed. The "content image" and "style image" are fed into the network for training. Different loss functions are constructed according to (1), (2), and (3) based on the output results according to features. Then, the inverse derivative is performed and the final simulated data with real noise is obtained through multiple iterations. (1) Content loss function: The input image and the generated image after style transfer are respectively, where l represents the l-th convolutional layer in the network, i represents the number of channels in the convolutional layer, and j represents the j-th channel in the convolutional layer. th At each position, F and P represent the feature responses of the corresponding outputs of the "content image" and "initial image" in the l-th layer; (2) Style loss function: First, we need to define the feature cross-correlation using the Gram matrix: The style loss function is: Here, 'a' represents the input style image, and 'M' represents the input style image. l Equals the height × width of a single feature map output by the l-th layer, N l The number of filters in this layer is represented by A and G, which are the cross-correlation between the feature responses of the "style image" and the corresponding features of the "initial image" in the l-th layer. l Represents the weights of the l-th layer; (3) Total loss function: Here, α and β are weighting factors that balance content loss and style loss.

Images