JP7442404B2 - Seismic waveform processing method and seismic waveform processing system - Google Patents

Description

The present invention relates to a technique for removing noise components from seismic observation data observed by a seismograph.

When an earthquake occurs, the earthquake is observed by seismometers installed at various locations, and seismic motion waveform data is generated as earthquake observation data. Earthquake observation data are used for various purposes, including estimating earthquake rupture processes.
This seismic observation data may contain noise from various vibrations that occur around the seismograph, such as road traffic vibrations, train vibrations, and construction vibrations. For example, in an emergency earthquake early warning system, etc., the arrival time of S waves is predicted from a small amount of information on P waves, but if earthquake observation data mixed with the noise mentioned above is used when constructing an emergency earthquake early warning system, etc. Errors may occur in the prediction of S waves. Therefore, efforts are being made to remove noise from seismic observation data.

For example, in Patent Document 1, vibrations measured by a seismograph are decomposed into intensity levels corresponding to frequency components, and a predetermined frequency of the vibrations is set to any frequency within the range of 15 to 25 (Hertz), for example, 20 (Hertz). ), and when the vibration level of the frequency component is a20 (decibel), the vibration noise level is an (decibel), and the vibration level at the vibration frequency fp (hertz) that gives the vibration peak is ap (decibel), r20 An earthquake observation method is disclosed in which when r20 (percent), given by = (a20-an)/(apan) x 100, exceeds a predetermined value, for example 60 (percent), it is determined to be noise and the vibration is removed. has been done.
In addition, Patent Document 2 describes the characteristics of seismic waves and their attributes using the calculation of the dominant frequency that is easily calculated using only integer operations, parameters for expressing how seismic waves rise, and the relationship between the two. An analysis method has been disclosed that determines that earthquake observation data that is not available is noise.
Further, in Patent Document 3, in a seismometer noise identification method, a surface seismometer that detects vibrations other than earthquakes that propagate on the ground surface, and an underground seismometer that detects seismic motion that propagates from underground to the ground surface are installed. However, if only the surface seismometer is activated, the vibration is identified as non-earthquake, and if both the underground seismometer and surface seismometer 4 are activated, the seismometer is identified as seismic vibration. A noise identification method is disclosed.

Noise identification and removal as described in the above-mentioned patent documents may not provide sufficient accuracy. Therefore, in noise processing of seismic observation data, a worker generally interprets the noise by visually observing the seismic observation data, and applies filter processing to the periodic band of the portion interpreted as noise.
However, at present, a huge number of earthquake records are made public by various public institutions, and it is not realistic to manually remove noise from all of these seismic observation data.
It is desired to automatically remove noise components with high precision.

Japanese Patent Application Publication No. 10-253766 Japanese Patent Application Publication No. 2009-103672 JP2014-66571A

The problem to be solved by the present invention is to provide a seismic waveform processing method and a seismic waveform processing system that can remove noise components with high accuracy.

In order to solve the above problems, the present invention employs the following means. That is, the present invention is a seismic waveform processing method for extracting seismic waveforms by removing noise components from seismic observation data observed by a seismograph. The images of the acceleration waveform and the velocity waveform of the seismic observation data are input to a first learning model that has been deep learned, using the information on the presence or absence of the noise component in the image as training data, and the information on the presence or absence of the noise component is determined by inputting the images of the acceleration waveform and the velocity waveform of the earthquake observation data. and when it is inferred that the noise component exists, an image in which a portion of the image containing the noise component has been identified is used as training data for input data including an image of an acceleration Fourier spectrum of an earthquake. inputting the image of the acceleration Fourier spectrum of the seismic observation data into a second deep learning model, inferring the noise component in the image and calculating the frequency boundary value of the noise component; A seismic waveform processing method is provided, comprising the step of generating a noise component-free seismic waveform in which the noise component is removed by filtering using the frequency boundary value.
According to the above configuration, the first learning model performs deep learning on input data including images of acceleration waveforms and velocity waveforms of an earthquake, using information on the presence or absence of noise components in the images as training data. There is. Therefore, images of the acceleration waveform and velocity waveform of the earthquake and information regarding the presence or absence of noise components in the images are learned as feature quantities in the first learning model. Therefore, when an image of an acceleration waveform and a velocity waveform of seismic observation data is input, the first learning model can appropriately infer whether or not there is a noise component in the image.
Further, the second learning model performs deep learning on input data including an image of an acceleration Fourier spectrum of an earthquake, using an image in which a portion containing a noise component in the image has been identified as training data. For this reason, the second learning model is trained with an image of the acceleration Fourier spectrum of the earthquake and information regarding a portion of the image that includes a noise component as feature quantities. Therefore, when an image of the acceleration Fourier spectrum of earthquake observation data is input, the second learning model determines which part of the image contains the noise component, that is, the frequency band in which the noise component of the acceleration Fourier spectrum is included. , can be properly reasoned.
In this way, the frequency band in which the noise component is included is inferred by the first and second learning models, so by calculating the boundary of this frequency band as a frequency boundary value and performing filter processing using this, It is possible to generate noise-free seismic waveforms with noise components removed. Therefore, it is possible to remove noise components with high accuracy.

In one aspect of the present invention, the seismic waveform processing method uses input data that includes the seismic waveform without noise components and a waveform in which sections of P waves and S waves are extracted by applying a filter to the seismic waveform. On the other hand, the third learning model that is deep learned using the arrival times of P waves and S waves as training data is either the seismic waveform of the earthquake observation data determined to be free of the noise component, or the seismic waveform from which the noise component has been removed. and inferring the arrival times of P waves and S waves by inputting either one of the following.
According to the above configuration, the third learning model uses input data that includes an earthquake waveform without noise components and a waveform in which P wave and S wave sections are extracted by applying a filter to the earthquake waveform. On the other hand, deep learning is performed using the arrival times of P waves and S waves as training data. Therefore, the third learning model includes an earthquake waveform without noise components, a waveform in which each section of P waves and S waves is extracted by applying a filter to the earthquake waveform, and the arrival times of P waves and S waves. The information about this is learned as a feature quantity. Therefore, when the third learning model receives either a seismic waveform of seismic observation data determined to have no noise component or a seismic waveform from which noise components have been removed as a seismic waveform with no noise component, the third learning model receives P waves and The arrival time of the S wave can be appropriately inferred.

The present invention also provides a seismic waveform processing system that extracts seismic waveforms by removing noise components from seismic observation data observed by a seismograph. Inputting images of the acceleration waveform and the velocity waveform of the earthquake observation data into a first learning model that has undergone deep learning using information on the presence or absence of the noise component in the image as training data, and the first learning model. and a noise component presence/absence inference unit that infers the presence or absence of the noise component, and an image in which a portion including the noise component in the image is identified as training data for input data including an image of an acceleration Fourier spectrum of an earthquake. , when it is inferred that there is a noise component, input the image of the acceleration Fourier spectrum of the seismic observation data to the second learning model that has undergone deep learning, and calculate the noise component in the image. a noise component removal unit that infers the noise component, calculates a frequency boundary value of the noise component, and generates a noise component-removed seismic waveform from which the noise component is removed by filtering using the frequency boundary value; A seismic waveform processing system is provided.
According to the above configuration, as already explained, it is possible to remove noise components with high accuracy.

According to the present invention, it is possible to provide a seismic waveform processing method and a seismic waveform processing system that can remove noise components with high accuracy.

FIG. 1 is a block diagram of a seismic waveform processing system in an embodiment of the present invention. FIG. 3 is an explanatory diagram of input data input to the first learning model of the seismic waveform processing system when the data does not include noise components. FIG. 3 is an explanatory diagram of input data input to the first learning model when the data includes a noise component. FIG. 3 is a schematic block diagram of a first learning model. FIG. 3 is a schematic block diagram of a second learning model of the seismic waveform processing system. FIG. 3 is an explanatory diagram of input data of a third learning model of the seismic waveform processing system. FIG. 7 is an explanatory diagram of teacher data of the third learning model. 3 is a flowchart of a seismic waveform processing method according to an embodiment. It is a graph showing the evaluation result of the third learning model.

The present invention is a seismic waveform processing method and seismic waveform processing system that removes noise components from seismic observation data and obtains P-wave and S-wave seismic waveforms using a machine learning model (AI technology).
Embodiments of the present invention will be described in detail below with reference to the drawings.
FIG. 1 is a block diagram of the seismic waveform processing system in this embodiment. The seismic waveform processing system 1 of this embodiment removes noise components from seismic observation data observed by a seismograph, and extracts seismic waveforms that do not include noise components. Furthermore, the seismic waveform processing system 1 infers the arrival times of P waves and S waves based on the seismic waveform from which noise components have been removed.
As shown in FIG. 1, the seismic waveform processing system 1 consists of a computer device, which removes noise components and extracts P waves and S waves by executing processing based on a preset computer program. Infer arrival time. The seismic waveform processing system 1 includes a noise component presence/absence inference section 2, a noise component removal section 3, and an arrival time derivation section 4. The noise component presence/absence inference section 2 includes a velocity waveform generation section 21, an input image generation section 22, and a first learning model 50. The noise component removal section 3 includes a spectrum generation section 31, a second learning model 60, a frequency boundary value extraction section 32, and a filter processing section 33. The arrival time derivation unit 4 includes an input data generation unit 41 and a third learning model 70.

The noise component presence/absence inference unit 2 inputs the acceleration waveform 5 of the earthquake observation data and infers whether or not the acceleration waveform 5 includes a noise component.
First, when the velocity waveform generation unit 21 receives the acceleration waveform 5 observed by the seismograph, it differentiates the acceleration waveform 5 and generates a velocity waveform.
Based on the acceleration waveform 5 and the velocity waveform generated by the velocity waveform generation unit 21, the input image generation unit 22 generates an image of an acceleration waveform and a velocity waveform (hereinafter referred to as a first image), which becomes input data to the first learning model 50. 1 input image) is generated. 2 and 3 are explanatory diagrams of the first input image 7. The input image generation unit 22 generates an acceleration waveform image 7a based on the acceleration waveform 5, generates a velocity waveform image 7b based on the velocity waveform, and further generates an acceleration waveform image 7a and a velocity waveform image 7b. are connected vertically to generate one first input image 7.

The first input image 7 shown in FIG. 2 is a waveform when the acceleration waveform 5 does not include a noise component. On the other hand, the first input image 7 shown in FIG. 3 contains noise components, and therefore features different from those in FIG. 2 appear in the image. The first learning model 50, which will be described next, has been machine learned to recognize this feature.
FIG. 4 is a schematic block diagram of the first learning model 50. In this embodiment, the first learning model 50 is realized by a convolutional neural network (hereinafter referred to as CNN) that is compatible with processing when images are used as input and output. The first learning model 50 includes a convolution processing section 51 and a full connection section 52.

The convolution processing unit 51 includes a plurality of convolution layers 51a. When the first input image 7 is input to the first learning model 50, convolution filter processing is performed by each of the plurality of filters provided in the first stage convolution layer 51a, and pooling processing etc. are applied as necessary. Then, a feature map corresponding to the first convolutional layer 51a is generated. By repeating this process to subsequent stages, a feature amount map 53 corresponding to the final stage convolution layer 51a is generated. This feature map 53 corresponding to the final convolutional layer 51a is input to the full combination unit 52.
The full coupling unit 52 includes an input layer 52a, a plurality of intermediate layers 52b, and an output layer 52c. Each of the values included in the input feature map 53 is input to a corresponding input node of the input layer 52a. A weighted sum is calculated for this value, and the result is stored in the first intermediate layer 52b. By repeating this process to subsequent stages, the final calculated value is stored in the output layer 52c.

The first learning model 50 solves two-class classification to infer whether or not the input first input image 7 contains a noise component. In this embodiment, the output layer 52c includes a first output node 52d and a second output node 52e. The first output node 52d corresponds to a case where the first input image 7 contains a noise component, and the second output node 52e corresponds to a case where it does not contain a noise component.
During learning, the first input image 7, which is input to the first learning model 50 as input data during learning, is processed by the convolution processing unit 51 and the full combination unit 52, and the first and second output nodes 52d, The processing result is stored in 52e. This processing result is compared with teacher data. In this embodiment, the teacher data is information on the presence or absence of noise components in the first input image 7. More specifically, for example, when the input first input image 7 contains noise, the processing result in the first output node 52d is 1, and the processing result in the second output node 52e is 0. be compared. Further, for example, when the input first input image 7 does not include noise, the processing result in the first output node 52d is compared with 0, and the processing result in the second output node 52e is compared with 1. Ru.
The first learning model 50 uses the error backpropagation method, the stochastic gradient descent method, etc. to reduce the cost function expressed by the squared error between these processing results and the teacher data. Machine learning is performed by adjusting the weight values of each filter, the weight values used for the weighted sum of the full combination unit 52, and the like.
As a result, in the first learning model 50, the values of the first output node 52d and the second output node 52e are close to 1 and 0, respectively, if the input first input image 7 contains a noise component. If not, the values are learned to be close to 0 and 1, respectively. In this way, the first learning model 50 is a trained learning model in which appropriate learning parameters have been learned, and is used as a program module that is part of artificial intelligence software.

The noise component presence/absence inference unit 2 inputs the first input image 7 to be inferred, generated by the input image generation unit 22, to the first learning model 50 trained as described above, and performs the first learning. The processing results stored in the first and second output nodes 52d and 52e of the model 50 are referred to. If the processing result stored in the first output node 52d is larger than the processing result stored in the second output node 52e and is closer to 1, the noise component presence/absence inference unit 2 determines that the first input image 7 contains a noise component. It is inferred that Conversely, if the processing result stored in the second output node 52e is larger and closer to 1 than the processing result stored in the first output node 52d, the noise component presence/absence inference section 2 adds a noise component to the first input image 7. It is inferred that it is not included.
In this way, the noise component presence/absence inference unit 2 inputs the first input image 7 of the earthquake observation data to the first learning model 50, infers the presence or absence of a noise component in the acceleration waveform 5, and infers the presence/absence of a noise component in the acceleration waveform 5. Send to.

When the noise component presence/absence inference unit 2 infers that the acceleration waveform 5 has a noise component, the noise component removal unit 3 inputs the acceleration waveform 5 and generates a noise component-removed seismic waveform from which the noise component is removed. do. FIG. 5 is a schematic block diagram of the noise component removal unit 3, particularly a second learning model 60 that will be described later.
First, the spectrum generation unit 31 receives the acceleration waveform 5 when the noise component presence/absence inference unit 2 infers that the acceleration waveform 5 includes a noise component. The spectrum generation unit 31 performs a Fourier transform on the acceleration waveform 5, and generates a Fourier spectrum image (hereinafter referred to as (referred to as the second input image) 8 is generated.
The spectrum generation unit 31 transmits and inputs the second input image 8 to the second learning model 60.

The second learning model 60 is subjected to machine learning so as to recognize which part of the input second input image 8 is a noise component as a feature.
In this embodiment, the second learning model 60 is realized by CNN, and is trained by semantic segmentation to infer which pixel of the second input image 8 corresponds to a noise component. There is. The second learning model 60 includes a convolution processing section 61 and a deconvolution processing section 62.
The convolution processing unit 61 includes a plurality of convolution layers 61a. When the second input image 8 is input to the second learning model 60, convolution processing is performed using the same procedure as the convolution processing unit 51 of the first learning model 50. The processing result of the convolution layer 61a at the final stage is a state in which the features of the second input image 8 are extracted and compressed to a low dimension, and this is input to the deconvolution processing section 62.
The deconvolution processing unit 62 executes deconvolution processing to expand and restore the features compressed to a low dimension by the convolution processing unit 61. The deconvolution processing unit 62 outputs an output image 64 as the output of the second learning model 60.

During learning, the second input image 8, which is input to the second learning model 60 as input data during learning, is processed by the convolution processing section 61 and the deconvolution processing section 62, and an output image 64 is output. This processing result is compared with a teacher image 65 which is teacher data. In this embodiment, the teacher image 65 is information on which pixels of the second input image 8 correspond to noise components, and is an image in which the portions of the second input image 8 that include noise components have been identified. be. More specifically, each pixel of the teacher image 65 is assigned a predetermined first pixel value when the corresponding pixel of the second input image 8 is a noise portion, and is assigned a predetermined first pixel value when the corresponding pixel of the second input image 8 is not a noise portion. It is set to have a predetermined second pixel value different from the pixel value. The output image 64 is generated by comparing the pixel values of each pixel with the teacher image 65 corresponding to the second input image 8 that has been input, and calculating, for example, the squared error of these differences between each pixel as a cost function. be done.
Then, the second learning model 60 is subjected to machine learning by adjusting the weight values of each filter using an error back propagation method or the like so as to reduce this cost function.
As a result, the second learning model 60 has a pixel value of each pixel of the output image 64 that is close to the first pixel value if the corresponding pixel of the input second input image 8 is a noise component; If it is not a component, the pixel value is learned to be close to the second pixel value. In this way, the second learning model 60 is a trained learning model that is used as a program module that is part of artificial intelligence software and has appropriate learning parameters learned.

The noise component removal unit 3 inputs the second input image 8 to be inferred, generated by the spectrum generation unit 31, to the second learning model 60 trained as described above. The second learning model 60 infers noise components in the second input image 8 and generates an output image 64.
The frequency boundary value extraction unit 32 receives the output image 64 and extracts a frequency band having the first pixel value within the output image 64. In particular, in this embodiment, the frequency boundary value extraction unit 32 extracts a low frequency band upper limit value that is the upper limit value of noise components that appear in the low frequency band, and a high frequency band lower limit value that is the lower limit value of the noise component that appears in the high frequency band. The value is calculated as the frequency boundary value of the noise component.
The filter processing unit 33 performs filter processing on the acceleration waveform 5 based on these frequency boundary values. More specifically, the filter processing unit 33 executes a high-pass filter on the acceleration waveform 5, setting the upper limit value of the low frequency band as the cutoff frequency and passing the frequency band higher than this, and setting the lower limit value of the high frequency band as the cutoff frequency. A cutoff frequency is set, and a low-pass filter is executed to pass the frequency band below this frequency. In this way, the filter processing unit 33 generates a noise component-free seismic waveform by removing noise components from the acceleration waveform 5.

The arrival time deriving unit 4 inputs either the seismic waveform of seismic observation data determined to have no noise component or the seismic waveform from which the noise component has been removed, and infers the arrival times of the P waves and S waves. That is, when the noise component presence/absence inference section 2 infers that the acceleration waveform 5 is free of noise, the arrival time derivation section 4 receives the acceleration waveform 5 as an input waveform. Furthermore, if it is inferred that there is noise, the arrival time deriving unit 4 receives as an input waveform the noise component removed seismic waveform from which the noise component has been removed by the noise component removing unit 3. In this way, in either case, the arrival time deriving unit 4 receives an earthquake waveform without noise components as an input waveform.

FIG. 6 is an explanatory diagram of input data of the third learning model. The input data generation unit 41 generates a data set 9, which is input data to the third learning model 70, which will be described next. In this embodiment, the data set 9 includes an acceleration waveform 91, an acceleration waveform 92 that saturates at 0.5 gal, and a deflection analysis result 93.
The acceleration waveform 91 includes an acceleration waveform 91a in the north-south direction, an acceleration waveform 91b in the east-west direction, and an acceleration waveform 91c in the vertical direction. The input data generation unit 41 receives input waveforms corresponding to each of the north-south direction, east-west direction, and up-down direction, normalizes each by dividing each of these by the maximum acceleration value, and generates each acceleration waveform 91a, 91b. , 91c (FIG. 6 shows the waveform before normalization).
An acceleration waveform 92 that saturates at 0.5 gal is an acceleration waveform 92a that saturates at 0.5 gal in the north-south direction, an acceleration waveform 92b that saturates at 0.5 gal in the east-west direction, and an acceleration waveform 92b that saturates at 0.5 gal in the vertical direction. The acceleration waveform 92c has an acceleration waveform 92c. The absolute value information of each of the acceleration waveforms 91a, 91b, and 91c is lost during the normalization described above. Each of the acceleration waveforms 92a, 92b, and 92c complements this and serves as an effective input for estimating the time when the P wave, which has a relatively small value, arrives.
The deflection analysis result 93 includes a P wave deflection analysis result 93a and an S wave deflection analysis result 93b. The deflection analysis results 93 are generated by performing Ross and Ben-Zion filter processing on the received input waveform to generate these deflection analysis results 93a and 93b. This filtering process determines the waveform section having the characteristics of P wave and S wave by deflection analysis of the observed seismic motion waveform of two horizontal components and one vertical component obtained at the observation point, and when this index is high, It can be determined that there is a high possibility that a P wave or an S wave has arrived.

The third learning model 70 is machine-learned to recognize features related to the P-wave arrival time and the S-wave arrival time in the input waveform. In this embodiment, the third learning model 70 is realized as a neural network having only fully connected parts without a convolution layer, as shown as the fully connected part 52 of the first learning model 50.
The third learning model 70 is given the above data set 9 as an input as a series of numerical data. Correspondingly, the input layer of the third learning model 70 includes a plurality of input nodes. When each data is input to the third learning model 70, by repeating the calculation of the weighted sum in the subsequent stage, the final calculation The value is stored.

The third learning model 70 infers the arrival times of P waves and S waves for the input data set 9. Correspondingly, the output layer of the third learning model 70 includes a plurality of output nodes corresponding to a series of numerical values related to the P-wave arrival time, and a plurality of output nodes corresponding to a series of numerical values related to the S-wave arrival time. It consists of output nodes.
More specifically, for example, when inferring the P wave arrival time with an accuracy of 0.1 seconds for an input waveform spanning 60 seconds, the P wave arrival time is 0.1 x 60 = 600. An output node corresponding to each time obtained by dividing 60 seconds into 0.1 second increments is provided. Similarly, regarding the arrival time of the S wave, for example, 600 output nodes are provided, and as a whole, for example, a total of 1200 output nodes are provided.
During learning, the data set 9 that is input to the third learning model 70 as input data during learning is processed by the full coupling unit, and the processing results are stored in each output node. This processing result is compared with teacher data. In this embodiment, the teacher data is a vector indicating the arrival time of the P wave and the arrival time of the S wave. FIG. 7 schematically shows a teacher vector 71, which is teacher data of the third learning model 70. The teacher vector 71 includes a P-wave teacher vector 71a and an S-wave teacher vector 71b. The P-wave teacher vector 71a is a one-hot vector in which each element is provided corresponding to each output node corresponding to the P-wave arrival time. That is, the P-wave teacher vector 71a has, for example, 600 elements as shown in the above example, and only the element corresponding to the time when the P-wave arrived in the input waveform is 1, and the element corresponding to other times is 1. All elements are expressed as 0. Similarly, the S-wave teacher vector 71b is a one-hot vector in which each element is provided corresponding to each output node corresponding to the S-wave arrival time. That is, the S-wave teacher vector 71b similarly has, for example, 600 elements, and only the element corresponding to the time when the S-wave arrives in the input waveform is 1, and all the elements corresponding to other times are 0. , is expressed. The processing results of the third learning model 70 are compared for each output node with the elements of the teacher vector 71 corresponding to each output node.
The third learning model 70 weights all connections using error backpropagation, stochastic gradient descent, etc. so as to reduce the cost function expressed by the squared error between these processing results and the training data. Machine learning is performed by adjusting the weight values used in the sum.
As a result, the third learning model 70 sets the value of the output node corresponding to each time inferred to be the P wave arrival time and the S wave arrival time in the input waveform to 1, and the values of other output nodes to 1. They are trained to be close to 0. In this way, the third learning model 70 is a trained learning model in which appropriate learning parameters have been learned, and is used as a program module that is part of artificial intelligence software.

The arrival time derivation unit 4 inputs the dataset 9 to be inferred, generated by the input data generation unit 41, to the third learning model 70 trained as described above, and calculates the third learning model 70. Refer to the processing results stored in each output node. The arrival time deriving unit 4 infers, as the arrival time of the P wave, the time corresponding to the output node having the largest value and closest to 1 among the output nodes corresponding to the P wave arrival time. Furthermore, the arrival time deriving unit 4 infers the time corresponding to the output node having the largest value and closest to 1 among the output nodes corresponding to the S wave arrival time, as the arrival time of the S wave.
The arrival time deriving unit 4 outputs the arrival times of the P waves and S waves inferred in this way as the arrival times 6.

Next, a seismic waveform processing method using the seismic waveform processing system 1 described above will be explained using FIGS. 1 to 7 and 8. FIG. 8 is a flowchart of the seismic waveform processing method.
When the process is started (step S1), the velocity waveform generation unit 21 acquires the acceleration waveform 5 observed by the seismograph, differentiates it, and generates a velocity waveform.
The input image generation unit 22 generates the first input image 7, which becomes input data to the first learning model 50, based on the acceleration waveform 5 and the velocity waveform generated by the velocity waveform generation unit 21 (step S3). .
The noise component presence/absence inference unit 2 inputs the first input image 7 of the earthquake observation data to the first learning model 50 and infers the presence or absence of a noise component (step S5).

Next, the seismic waveform processing system 1 determines whether or not there is a noise component in the acceleration waveform 5 based on the noise component inference result (step S7).
If it is determined that there is a noise component (Yes in step S7), the spectrum generation unit 31 performs Fourier transform on the acceleration waveform 5 to generate the second input image 8. The noise component removal unit 3 inputs the second input image 8 to the second learning model 60, infers the noise component in the second input image 8, and generates an output image 64. The frequency boundary value extraction unit 32 receives the output image 64, extracts a frequency band having the first pixel value within the output image 64, and calculates the frequency boundary value of the noise component (step S9). The filter processing unit 33 performs filter processing on the acceleration waveform 5 based on the frequency boundary value to generate a noise component removed seismic waveform (step S11). After that, the process transitions to step S13, which will be described next.
Even when it is determined that there is no noise component (No in step S7), the process moves to step S13.

In step S13, the arrival time deriving unit 4 determines that the seismic waveform (acceleration waveform 5) of the earthquake observation data is present if it is determined that there is no noise component in step S7, or that there is a noise component in step S7. If so, the noise component removed seismic waveform generated in step S11 is received, and based on this, the arrival times of the P waves and S waves are inferred.
More specifically, the input data generation unit 41 generates the data set 9 that is input data to the third learning model 70. The arrival time deriving unit 4 inputs this to the third learning model 70, infers the arrival time of the P wave and the S wave, and outputs it as the arrival time 6. After that, the process ends (step S15).

Next, the effects of the seismic waveform processing method and seismic waveform processing system described above will be explained.
The seismic waveform processing method in this embodiment is an seismic waveform processing method for extracting seismic waveforms by removing noise components from seismic observation data observed by a seismograph. The first learning model 50, which has undergone deep learning, uses the information on the presence or absence of noise components in the image 7 as training data for the input data including the input image) 7. A step of inputting the first input image) 7 and inferring the presence or absence of a noise component (step S5), and when it is inferred that there is a noise component, an image of the earthquake acceleration Fourier spectrum (second input image) 8 The second learning model 60 performs deep learning using an image (teacher image) 65 in which a portion of the image 8 containing a noise component has been identified as the teacher data. The image (second input image) 8 is input, the noise component in the image 8 is inferred, the frequency boundary value of the noise component is calculated, and the noise component is filtered by filtering using the frequency boundary value. The method includes a step of generating a seismic waveform from which noise components have been removed (steps S9 and S11).
Furthermore, the seismic waveform processing system 1 in this embodiment is an seismic waveform processing system 1 that removes noise components from seismic observation data observed by a seismograph to extract seismic waveforms. A first learning model 50 is deep trained on input data including an image (first input image) 7, using information on the presence or absence of noise components in the image 7 as training data, and the first learning model 50 is A noise component presence/absence inference unit 2 receives an image of the acceleration waveform and velocity waveform of observation data (first input image) 7 and infers the presence or absence of a noise component, and an image of the acceleration Fourier spectrum of the earthquake (second input image). 8, the second learning model 60 performs deep learning using an image (teacher image) 65 in which a portion of the image 8 containing a noise component has been identified as training data, and infers that there is a noise component. In this case, input the image (second input image) 8 of the acceleration Fourier spectrum of the earthquake observation data to the second learning model 60, infer the noise component in the image 8, and determine the frequency boundary of the noise component. and a noise component removal unit 3 that calculates the value and generates a noise component-free seismic waveform in which the noise component is removed by filtering using the frequency boundary value.
According to the above configuration, the first learning model 50 acquires information on the presence or absence of noise components in the image 7 for input data including an image (first input image) 7 of an acceleration waveform and velocity waveform of an earthquake. is used as training data for deep learning. For this reason, the first learning model 50 has learned an image (first input image) 7 of the acceleration waveform and velocity waveform of the earthquake, and information regarding the presence or absence of noise components in the image 7 as feature quantities. . Therefore, when an image (first input image) 7 of acceleration waveforms and velocity waveforms of earthquake observation data is input, the first learning model 50 appropriately infers whether or not there is a noise component in the image 7. can.
The second learning model 60 also uses an image (teacher image) 65 in which a portion of the image 8 containing noise components has been identified for input data including an image (second input image) 8 of an acceleration Fourier spectrum of an earthquake. is used as training data for deep learning. For this reason, the second learning model 60 has learned the image (second input image) 8 of the acceleration Fourier spectrum of the earthquake and information regarding the portion containing the noise component in the image 8 as feature quantities. Therefore, when an image (second input image) 8 of the acceleration Fourier spectrum of earthquake observation data is input, the second learning model 60 determines which part of the image 8 contains the noise component, that is, the acceleration Fourier spectrum. It is possible to appropriately infer the frequency band in which the noise component of .
In this way, the first and second learning models 50 and 60 infer the frequency band in which the noise component is included, so the boundary of this frequency band is calculated as a frequency boundary value, and filter processing is performed using this. By doing so, it is possible to generate a noise component-free seismic waveform with noise components removed. Therefore, in the machine learning demo model (first and second learning models 50, 60), for example, a threshold value is not set to remove noise components for a specific frequency band, and a threshold value is not necessary. It is possible to remove noise components with high accuracy.

In particular, in this embodiment, both the acceleration waveform and velocity waveform of an earthquake are used as input to the first learning model 50. Short-period noise components tend to appear in acceleration waveforms, and long-period noise components tend to appear in velocity waveforms, so by using both as input, noise components can be efficiently removed regardless of the periodic band of the noise components. can be detected.

In addition, the above seismic waveform processing method uses a seismic waveform with no noise component (acceleration waveform 5 when it is determined that there is no noise component, or a seismic waveform with noise component removed) and a filter applied to the seismic waveform. The input data including the waveform (deflection analysis result) 93 from which each section of the wave and the S wave has been extracted is subjected to deep learning using the arrival times of the P wave and the S wave as training data. The method includes a step of inputting either a seismic waveform of seismic observation data determined to have no component or a seismic waveform from which noise components have been removed, and inferring the arrival times 6 of P waves and S waves.
According to the above configuration, the third learning model 70 receives input data including a seismic waveform without noise components and a waveform in which sections of P waves and S waves are extracted by applying a filter to the seismic waveform. In contrast, deep learning is performed using the arrival times of P waves and S waves as training data. Therefore, the third learning model 70 includes an earthquake waveform without noise components, a waveform in which each section of P waves and S waves is extracted by applying a filter to the earthquake waveform, and arrival of P waves and S waves. Information regarding time is learned as a feature quantity. Therefore, when either the seismic waveform of seismic observation data determined to have no noise component or the seismic waveform from which the noise component has been removed is input, the third learning model 70 calculates the arrival times 6 of P waves and S waves. , can be properly reasoned.

FIG. 9 is a graph showing the evaluation results of the third learning model 70 constructed as described above. This is a histogram of the residual with the result inferred by . It can be seen that the results are concentrated near 0 seconds and the residual is close to 0, that is, the accuracy of the third learning model 70 is high.

The seismic waveform processing method and seismic waveform processing system of the present invention are not limited to the above-described embodiments described with reference to the drawings, and various other modifications can be considered within the technical scope thereof. .
For example, in the above embodiment, the first learning model 50 uses the image (first input image) 7 of the acceleration waveform and velocity waveform of an earthquake as input data, but in addition to this, other data is input and given. It's okay.
Further, in the above embodiment, the second learning model 60 uses the image (second input image) 8 of the acceleration Fourier spectrum of the earthquake as input data, but in addition to this, other data may also be input. good.
Furthermore, in the above embodiment, the third learning model 70 uses the data set 9 as input data, but this data set 9 includes an acceleration waveform 91, an acceleration waveform 92 that saturates at 0.5 gal, and a deflection analysis result 93. Other data may also be included.
In addition, when realizing the first, second, and third learning models 50, 60, and 70, it is not necessary to use the configuration described in the above embodiment, and for example, a machine learning device other than CNN may be used. Needless to say, it's okay.
In addition to this, it is possible to select the configurations mentioned in each of the above embodiments or to change them to other configurations as appropriate without departing from the gist of the present invention.

1 Seismic waveform processing system 5 Acceleration waveform 2 Noise component presence inference section 6 Arrival time 21 Velocity waveform generation section 7 First input image (image of acceleration waveform and velocity waveform)
22 Input image generation unit 8 Second input image (image of acceleration Fourier spectrum)
3 Noise component removal unit 50 First learning model 31 Spectrum generation unit 60 Second learning model 32 Frequency boundary value extraction unit 65 Teacher image (image in which a portion containing noise components has been identified)
33 Filter processing section 70 Third learning model 4 Arrival time derivation section 91 Earthquake waveform (acceleration waveform) without noise components
41 Input data generation unit 93 Deflection analysis result (waveform from which sections are extracted)
42 Arrival time extraction part

Claims (3)

A seismic waveform processing method for extracting seismic waveforms by removing noise components from seismic observation data observed by a seismograph, the method comprising:
The acceleration waveform of the earthquake observation data is applied to input data including images of the acceleration waveform and velocity waveform of an earthquake. and inputting an image of the speed waveform to infer the presence or absence of the noise component;
When it is inferred that the noise component exists, deep learning is performed on the input data including the image of the acceleration Fourier spectrum of the earthquake, using the image in which the part of the image containing the noise component has been identified as training data. The image of the acceleration Fourier spectrum of the seismic observation data is input to the second learning model, the noise component in the image is inferred, the frequency boundary value of the noise component is calculated, and the frequency boundary value is used. 1. A seismic waveform processing method, comprising the step of: generating a noise component-free seismic waveform in which the noise component is removed by filtering. The arrival times of P waves and S waves are used as training data for input data including the seismic waveform without noise components and a waveform in which each section of P waves and S waves is extracted by applying a filter to the seismic waveform. , either the seismic waveform of the seismic observation data determined to have no noise component or the seismic waveform from which the noise component has been removed is input into the deep learning third learning model, and the P-wave and S-wave The seismic waveform processing method according to claim 1, further comprising the step of inferring an arrival time. A seismic waveform processing system that extracts seismic waveforms by removing noise components from seismic observation data observed by a seismograph,
a first learning model that performs deep learning on input data including images of acceleration waveforms and velocity waveforms of an earthquake, using information on the presence or absence of the noise component in the image as training data;
a noise component presence/absence inference unit that inputs images of the acceleration waveform and the velocity waveform of the earthquake observation data into the first learning model and infers the presence or absence of the noise component;
a second learning model that performs deep learning on input data including an image of an acceleration Fourier spectrum of an earthquake, using an image in which the portion containing the noise component in the image has been identified as training data;
When it is inferred that the noise component exists, input the image of the acceleration Fourier spectrum of the seismic observation data to the second learning model, infer the noise component in the image, and calculate the noise component. A seismic waveform processing system comprising: a noise component removal unit that calculates a frequency boundary value and generates a noise component-free seismic waveform in which the noise component is removed through filter processing using the frequency boundary value.

Images