Earthquake first arrival picking method based on convolutional neural network

Abstract

The present invention discloses a method for picking up earthquake first arrivals based on a convolutional neural network to address the problems of inaccurate first arrival picking of pre-stack time domain seismic data and the need for repeated manual parameter adjustment. The method comprises: preprocessing seismic data; extracting high-dimensional features of seismic signals using a convolutional network; generating the distribution area of ​​the first arrival "before and after arrival" using a deconvolutional network; and determining the first arrival time of an earthquake. The present invention can automatically analyze the arrival time of the first arrival through a convolutional neural network, thereby improving the accuracy and intelligence of picking up the first arrival. The method solves the problem that the prior art uses an improved method based on the energy ratio of long and short time windows, which not only has a large deviation in picking up the first arrival under low and medium signal-to-noise ratio conditions, but also requires manual tuning of multiple parameters and has a high trial and error cost.

Description

Technical Field

[0001] The present invention relates to the field of petroleum geophysical exploration data processing, and in particular to a seismic first arrival picking method based on convolutional neural network. Background Art

[0002] In seismic exploration, the seismic wave that is excited from the shot point and first reaches the geophone is called the first arrival wave, and the arrival time is called the seismic first arrival time, or seismic first arrival for short. Seismic first arrival picking is directly related to the quality of static correction and near-surface velocity tomography inversion. Especially in complex surface exploration areas such as deserts and mountains, the accuracy of first arrival picking is directly related to the success or failure of seismic processing projects.

[0003] The most classic method for picking up earthquake first arrivals is the long-short time window average energy ratio method (STA / LTA). Its characteristic is that the first arrival position is calculated on a single-channel basis. Moreover, parameters such as the length of the time window and the threshold of the energy ratio need to be manually defined. This method and its many improved methods have good picking effects in high signal-to-noise ratio areas, but in medium and low signal-to-noise ratio areas, the application effect is generally unsatisfactory. The main manifestations are: (1) The automatic picking accuracy is less than 50%. (2) Manual parameter adjustment requires a lot of experiments, and it is difficult to find the optimal configuration.

[0004] With the development of artificial intelligence, especially deep convolutional neural network technology, breakthroughs have been made in image recognition, target detection, semantic segmentation and other tasks, and the accuracy of classification has been significantly improved. Therefore, it is urgent to develop a method for picking earthquake first arrivals based on convolutional neural networks. Summary of the invention

[0005] The purpose of the present invention is to solve the deficiencies in the prior art and to provide a seismic first-arrival picking method based on a convolutional neural network, which can replace manual picking or traditional methods and improve the accuracy and intelligence of first-arrival picking.

[0006] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0007] The method for picking earthquake first arrival based on convolutional neural network is characterized by comprising:

[0008] Step 101, preprocessing the seismic data;

[0009] Step 102, extracting high-dimensional features of seismic signals using a convolutional network;

[0010] Step 103, using a deconvolution network to generate the distribution areas of the first arrival "before arrival and after arrival";

[0011] Step 104, determine the first arrival time of the earthquake.

[0012] The above scheme further comprises:

[0013] The preprocessing includes: segmenting the seismic record file and normalizing the seismic signal attributes.

[0014] The segmentation of the seismic record file refers to segmenting the seismic record file into single shot records according to the trace header mark, and training and predicting the convolution network and the deconvolution network based on the single shot record as the granularity.

[0015] Assume that the amplitude value of the i-th sampling point in the j-th channel of a single shot record is A(i,j), then the seismic signal attributes include: A(i,j), |A(i,j)|, A(i,j) 2 、A(i,j)-A(i-1,j),A(i,j) 2 -A(i-1,j)A(i+1,j).

[0016] The normalization refers to standardizing or normalizing the seismic signal attributes of all sampling points within the range of a single shot record to form a seismic signal attribute matrix of the single shot record.

[0017] The convolutional network refers to a stacked deep convolutional network composed of multiple convolutional modules connected sequentially or jump-connected; each convolutional module contains one or more convolution, pooling, activation function, random inactivation, and group normalization layers; the input of the convolutional network is the normalized seismic signal attribute matrix of a single shot record.

[0018] The high-dimensional features refer to a multi-channel feature matrix obtained by performing nonlinear calculations on a single-channel feature matrix through a convolutional network.

[0019] The deconvolution network refers to a stacked deep deconvolution network composed of multiple deconvolution modules connected sequentially or jump-connected; each deconvolution module contains one or more deconvolution, activation function, and random inactivation layer; the last deconvolution module adopts a softmax binary classification activation function, and the output of the deconvolution network is a 0-1 classification matrix of a single shot record.

[0020] The 0-1 classification matrix of the single shot record means that each sampling point of the single shot record is calculated through the deconvolution network to obtain a classification result of 0 or 1; wherein 0 represents before the first arrival and 1 represents after the first arrival.

[0021] The distribution area of ​​the first arrival "before and after arrival" refers to the first arrival pre-arrival area composed of the set of sampling points classified as 0 in the 0-1 classification matrix output by the deconvolution network, and the first arrival post-arrival area composed of the set of sampling points classified as 1.

[0022] The convolution network and the deconvolution network are trained in the following manner: initializing the convolution kernel parameters of the convolution network and the deconvolution kernel parameters of the deconvolution network, inputting a normalized seismic signal attribute matrix, or multiple normalized seismic signal attribute matrices with the same rows and columns, into the convolution network and the deconvolution network to output a 0-1 classification matrix; quantifying the difference between the predicted "before arrival and after arrival" distribution area of ​​the first arrival and the manually marked distribution area through a cross entropy loss function, and updating the convolution kernel parameters of the convolution network and the deconvolution kernel parameters of the deconvolution network until the difference value calculated by the cross entropy loss function converges to a smaller value.

[0023] The first arrival time refers to first calculating the first arrival sampling point position of each column according to the 0-1 classification matrix output by the deconvolution network; then obtaining the sampling time interval in the seismic record file; and multiplying the two is the first arrival time.

[0024] Assume that the 0-1 classification matrix C has m rows and n columns, which means that the single shot record has n channels, each with m sampling points; then the first arrival sampling point position refers to: the position (i, j) in the column vector, which is the first sampling point position in the column vector that is determined to be classified as 1, and is defined as follows:

[0025] 0<i≤m,0<j≤n

[0026] 0<k<i,C(k,j)=0

[0027] C(i,j)=1

[0028] Or it is the starting sampling point position of multiple consecutive sampling positions in the column vector that are determined to be classified as 1, defined as follows:

[0029]

[0030]

[0031]

[0032] Where N means that all N consecutive sampling points belong to category 1. Assuming that the first arrival is at sampling point (i, j), k represents the number of the sampling point before the i-th sampling point in channel j. w represents the number of the N consecutive sampling points from sampling point (k, j) to sampling point (k+N-1, j).

[0033] Compared with the prior art, the present invention can automatically distinguish the areas before and after the first arrival through a convolutional neural network and accurately calculate the first arrival time. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of a method for picking earthquake first arrival based on convolutional neural network provided by an example of the present invention.

[0035] Figure 2 It is an exemplary structural diagram of a convolutional network and a deconvolutional network provided by an example of the present invention.

[0036] Figure 3 This is a rendering of the distribution area before and after arrival provided by an example of the present invention.

[0037] Figure 4 It is a schematic diagram of calculating the first arrival time of an earthquake provided by an example of the present invention. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are only used to explain the present invention and are not used to limit the invention.

[0039] Embodiment 1

[0040] The seismic first arrival picking method based on convolutional neural network includes:

[0041] Step 101, preprocessing the seismic data;

[0042] Step 102, extracting high-dimensional features of seismic signals using a convolutional network;

[0043] Step 103, using a deconvolution network to generate the distribution areas of the first arrival "before arrival and after arrival".

[0044] Step 104, determine the first arrival time of the earthquake.

[0045] Furthermore, the preprocessing in 101 includes: segmenting the seismic record file and normalizing the seismic signal attributes. Segmenting the seismic record file means segmenting the seismic record file into single-shot records according to the trace header mark, and training and predicting the convolution network and the deconvolution network with the single-shot record as the granularity. Assuming that the amplitude value of the j-th trace and the i-th sampling point of the single-shot record is A(i,j), the seismic signal attributes include: A(i,j), |A(i,j)|, A(i,j) 2 、A(i,j)-A(i-1,j),A(i,j) 2 -A(i-1,j)A(i+1,j). Normalization refers to standardizing or normalizing the seismic signal attributes of all sampling points within the range of a single shot record to form a seismic signal attribute matrix of a single shot record.

[0046] Furthermore, the convolutional network in 102 refers to a deep convolutional network composed of multiple convolutional modules connected sequentially or jump-connected. Each convolutional module contains one or more convolution, pooling, activation function, random inactivation, and group normalization layers. The input of the convolutional network is the normalized seismic signal attribute matrix recorded by a single shot. High-dimensional features refer to the nonlinear calculation of a single-channel feature matrix through a convolutional network to obtain a multi-channel feature matrix.

[0047] Furthermore, the deconvolution network in 103 refers to a deep deconvolution network composed of multiple deconvolution modules connected sequentially or jump-connected. Each deconvolution module contains one or more deconvolution, activation function, and random inactivation layer. The last deconvolution module adopts a softmax binary classification activation function, and the output of the deconvolution network is a 0-1 classification matrix of a single shot record. Specifically, 0 represents before the first arrival, and 1 represents after the first arrival. The area before the first arrival is composed of the set of all sampling points classified as 0, and the area after the first arrival is composed of the set of all sampling points classified as 1.

[0048] Furthermore, the convolution network in 102 and the deconvolution network in 103 are trained in the following manner: the convolution kernel parameters of the convolution network and the deconvolution kernel parameters of the deconvolution network are initialized, a normalized seismic signal attribute matrix, or multiple normalized seismic signal attribute matrices with the same rows and columns, are input into the convolution network, and the deconvolution network outputs a 0-1 classification matrix. The difference between the distribution area of ​​the predicted first arrival "before arrival and after arrival" and the distribution area of ​​the manual marking is quantified by the cross entropy loss function, and the convolution kernel parameters of the convolution network and the deconvolution kernel parameters of the deconvolution network are updated until the difference value calculated by the cross entropy loss function converges to a smaller value.

[0049] Furthermore, the first arrival time in step 104 refers to the calculation of the first arrival sampling point position of each channel. Assume that the 0-1 classification matrix C output by the deconvolution network has a total of m rows and n columns, which means that the single shot record has n channels, each with m sampling points. Then the first arrival sampling point position refers to: the position (i, j) in the column vector, which is the first sampling point position determined as 1 classification in the column vector, and is defined as follows:

[0050] 0<i≤m,0<j≤n

[0051] 0<k<i,C(k,j)=0

[0052] C(i,j)=1

[0053] Or it is the starting sampling point position of multiple consecutive sampling positions in the column vector that are determined to be classified as 1, defined as follows:

[0054]

[0055]

[0056]

[0057] Where N represents the length of continuous sampling points.

[0058] Then get the sampling time interval in the seismic record file, assumed to be span.

[0059] Then the first arrival time is the product of the two, that is, i*span.

[0060] Embodiment 2

[0061] Reference Figure 1 , the flowchart of the seismic first arrival picking method based on convolutional neural network includes steps 101-104. In step 101, the seismic data is preprocessed. As an example, the seismic data can be saved in a file in segy format and divided into single shot records according to the header mark. Assuming that a single shot record has n channels and each channel has m sampling points, the input matrix shape of the convolutional network is (m,n). Assuming that the amplitude value of the j-th channel and the i-th sampling point is A(i,j), then the input matrix value can select the following seismic signal attributes: A(i,j), |A(i,j)|, A(i,j) 2 、A(i,j)-A(i-1,j),A(i,j) 2 -A(i-1,j)A(i+1,j). Then normalize or standardize the values ​​of the input matrix. In addition, the convolutional network generally requires the input matrix to have the same shape. Considering the deconvolution operation, both m and n must be multiples of 2. However, it is very likely that the number of gathers or sampling points in a single shot record is different. In this case, the data can be cropped or filled as needed.

[0062] The present invention uses single shot records as the granularity to train and predict the convolutional network and the deconvolutional network. Other methods need to randomly or sequentially crop single shot records into fixed-size blocks such as 64*64 or 128*128, which will lead to very inaccurate first-arrival region classification.

[0063] In step 102, a convolutional network is used to extract high-dimensional features of the seismic signal. The convolutional network can sequentially connect or jump-connect multiple convolutional modules to form a stacked deep convolutional network. Each convolutional module contains one or more convolution, pooling, activation function, random inactivation, and group normalization layers. The input of the convolutional network is a normalized seismic signal attribute matrix, and the output is a high-dimensional feature of the initial distribution extracted by the convolutional network. As an example, the convolution kernel generally uses 1*1, 3*3, 5*5 rules, and the moving step size is a multiple of 2. Pooling generally selects the maximum value in the 2*2 rule. The activation function can use nonlinear activation functions such as RELU, LeakyRELU, ELU, etc.

[0064] In step 103, the distribution area of ​​the first arrival "before arrival and after arrival" is generated by the deconvolution network. The deconvolution network can sequentially connect or jump connect multiple deconvolution modules to form a stacked deep deconvolution network. Each deconvolution module contains one or more deconvolution, activation function, and random inactivation layer. The last deconvolution module adopts the softmax binary classification activation function. The output of the deconvolution network is a 0-1 classification matrix recorded by a single shot, and the shape of the matrix must be consistent with the shape of the input matrix, that is, (m, n). Specifically, 0 represents before the first arrival, and 1 represents after the first arrival. As an example, the deconvolution kernel generally uses the rules of 1*1, 3*3, and 5*5, and the moving step size is selected as a multiple of 2. Except for the last layer, the activation function can use nonlinear activation functions such as RELU, LeakyRELU, and ELU.

[0065] In step 104, the first arrival sampling point position of each column is calculated. Assume that the 0-1 classification matrix C output by the deconvolution network has a total of m rows and n columns, which means that the single shot record has n channels, each with m sampling points. Then the first arrival sampling point position refers to: the position (i, j) in the column vector, which is the first sampling point position in the column vector that is determined to be classified as 1, and is defined as follows:

[0066] 0<i≤m,0<j≤n

[0067] 0<k<i,C(k,j)=0

[0068] C(i,j)=1

[0069] Or it is the starting sampling point position of multiple consecutive sampling positions in the column vector that are determined to be classified as 1, defined as follows:

[0070]

[0071]

[0072]

[0073] Where N means that all N consecutive sampling points belong to category 1. Assuming that the first arrival is at sampling point (i, j), k represents the number of the sampling point before the i-th sampling point in channel j. w represents the number of the N consecutive sampling points from sampling point (k, j) to sampling point (k+N-1, j).

[0074] Then get the sampling time interval in the seismic record file, assumed to be span.

[0075] Then the first arrival time is the product of the two, that is, i*span.

[0076] Embodiment 3

[0077] Reference Figure 2 , the exemplary structural diagram of the convolutional network and the deconvolutional network in the earthquake first arrival picking method requires steps 201-205. As an example, the convolutional network structure has 4 layers, including 3 convolutional modules, which are connected sequentially, corresponding to steps 201-203. The convolutional network structure has 2 layers, including 2 convolutional modules, which are connected sequentially, corresponding to steps 204-205.

[0078] As an example, a single shot record has 128 channels, each with 800 sampling points, and is a single channel, so the normalized seismic signal attribute matrix of the first layer of the convolutional network has a shape of (128,800,1). In step 201, the convolution module includes 32 convolution and activation functions and 1 pooling function for feature extraction. The convolution uses a 3*3 convolution kernel, the activation function uses RELU, and the pooling uses the maximum value of the 2*2 rule. Therefore, the feature matrix of the second layer of the convolutional network has a shape of (64,400,32), forming a multi-channel matrix.

[0079] In step 202, the convolution module includes 32 convolution and activation functions, 1 pooling function, and 1 random dropout layer for feature extraction. The convolution uses a 3*3 convolution kernel, the activation function uses RELU, the maximum pooling of 2*2, and the dropout ratio is set to 0.5. Therefore, the feature matrix of the third layer of the convolution network has a shape of (32, 200, 32), forming a multi-channel matrix.

[0080] In step 203, the convolution module includes 256 convolution and activation functions, 1 pooling function, and 1 random dropout layer for feature extraction. The convolution uses a 1*1 convolution kernel, the activation function uses RELU, the maximum pooling of 2*2, and the dropout ratio is set to 0.5. Therefore, the feature matrix of the fourth layer of the convolution network has a shape of (16, 100, 256), which is the high-dimensional feature that describes the initial position.

[0081] In step 204, the deconvolution module includes 32 convolution and activation functions, 1 pooling function, and 1 random dropout layer. The convolution uses a 1*1 convolution kernel, the activation function uses RELU, the maximum pooling of 2*2, and the dropout ratio is set to 0.5. Therefore, the first layer feature matrix of the deconvolution network has a shape of (64,400,32).

[0082] In step 205, the deconvolution module includes 2 convolution and activation functions and 1 pooling function. The convolution uses a 2*2 convolution kernel, the activation function uses softmax, and the maximum value pooling of 2*2. Therefore, the second-layer feature matrix of the deconvolution network has a shape of (128,800,2). The third dimension of the matrix generates the probability values ​​of classification as 0 and classification as 1 at each sampling point. If the value classified as 1 is greater than the value classified as 0, it is considered that the sampling point should be defined as classification 1, that is, it belongs to the area after the first arrival. In this way, the "before arrival and after arrival" distribution map of the first arrival can be generated.

[0083] This example only shows the case where the convolution module of the convolutional network and the deconvolution module of the deconvolution network are connected sequentially. If the initial recognition of the "before arrival and after arrival" areas is inaccurate, you can consider optimizing by skipping feature matrices of the same shape, or increasing the number of network layers and convolution kernels.

[0084] Embodiment 4

[0085] Reference Figure 3 , generating the effect map of the "before arrival and after arrival" area requires steps 301-304. In step 301, a segmented single shot record is imaged by commercial seismic processing software. Figure 3 After normalizing the seismic signal attributes, the input matrix of the convolutional network is obtained.

[0086] In step 302, high-dimensional features of the first arrival distribution are extracted through a convolutional network.

[0087] In step 303, the high-dimensional features of the first-arrival distribution are mapped into a 0-1 classification matrix through a deconvolution network. The matrix can be considered to classify all the sampling points recorded in a single shot.

[0088] In step 304, the 0-1 classification matrix is ​​imaged as Figure 3 , where the yellow area represents the set of sampling points classified as 1, and the purple area represents the set of sampling points classified as 0. In other words, the yellow area is the distribution area after the first arrival, and then the position of the first arrival sampling point of each channel can be determined.

[0089] Embodiment 5

[0090] Reference Figure 4 , steps 401-404 are required to calculate the earthquake first arrival time diagram. In step 401, the 0-1 classification matrix output by the deconvolution network is obtained, which is recorded as C. As an example, C has 16 rows and 10 columns, which means that the single shot record has 10 channels, each with 16 sampling points. Among them, 0 represents before the first arrival, and 1 represents after the first arrival.

[0091] In step 402, the position of the first arrival sampling point is calculated. This example will illustrate two methods: (1) The position of the first sampling point in the column vector that is determined to be classified as 1 is defined as follows:

[0092] 0<i≤m,0<j≤n

[0093] 0<k<i,C(k,j)=0

[0094] C(i,j)=1

[0095] In this example, the first column vector [0,0,1,0,0,0,0,0,0,0,0,0,1,1,1,1,1,1] has an initial arrival position of i=3, j=1, indicating that the third sampling point of the first track is the initial arrival position. Similarly, we can get the initial arrival position vector [3, 7, 5, 5, 4, 2, 7, 10, 10, 15] corresponding to C, corresponding to the initial arrival position of each track.

[0096] (2) Multiple consecutive sampling positions in the column vector are determined as the starting sampling point positions of the 1st classification, which are defined as follows:

[0097]

[0098]

[0099]

[0100] In this example, assuming that N=2 is defined, the first column vector [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1] has an initial position of i=12, j=1, indicating that the 12th sampling point of the first channel is the initial position. Similarly, the initial position vector corresponding to C can be obtained as [12, 7, 5, 5, 4, 2, 7, 10, 13, 15], corresponding to the initial position of each channel. Obviously, method (2) has better fault tolerance than method (1).

[0101] In step 403, the sampling interval is recorded in the first 400 bytes of the binary file header in the segy file. As an example, assume that the sampling interval of the segy file where the single shot record is located is 2000 microseconds.

[0102] In step 404, the product of the first arrival position vector and the sampling interval is the first arrival time. For example, if the first arrival position vector is [3, 7, 5, 5, 4, 2, 7, 10, 10, 15] and the sampling interval is 2000 microseconds, then the first arrival time vector is:

[0103] [3, 7, 5, 5, 4, 2, 7, 10, 10, 15]*2000=[6000, 14000, 10000, 10000, 8000, 4000, 14000, 20000, 20000, 330000]

[0104] Corresponding to the first arrival time of each track.

Claims

1. A method for picking earthquake first arrivals based on convolutional neural networks, characterized in that: include: Step 101, preprocessing the seismic data; Step 102, using a convolutional network to extract high-dimensional features of the seismic signal; Step 103, using a deconvolution network to generate the distribution areas of the first arrival "before arrival and after arrival"; Step 104, determine the first arrival time of the earthquake; the preprocessing of the seismic data includes: segmenting the seismic record file and normalizing the seismic signal attributes; the segmentation of the seismic record file means that the seismic record file is segmented into single shot records according to the trace header mark, and the convolution network and the deconvolution network are trained and predicted with the single shot record as the granularity; assuming that the amplitude value of the j-th trace and the i-th sampling point of the single shot record is A(i,j), then the seismic signal attributes include: A(i,j), |A(i,j)|, A(i,j) 2 、A(i,j)-A(i-1,j),A(i,j) 2 -A(i-1,j)A(i+1,j); the normalization refers to standardizing or normalizing the seismic signal attributes of all sampling points within the range of a single shot record to form a seismic signal attribute matrix of a single shot record; the convolutional network refers to a stacked deep convolutional network composed of multiple convolutional modules connected in sequence or in jumps; each convolutional module contains one or more convolution, pooling, activation function, random inactivation, and group normalization layers; the input of the convolutional network is the normalized seismic signal attribute matrix of a single shot record; the high-dimensional feature refers to a single-channel feature matrix obtained by nonlinear calculation of the convolutional network to obtain a multi-channel feature matrix; the deconvolutional network refers to a stacked deep deconvolutional network composed of multiple deconvolutional modules connected in sequence or in jumps; each deconvolutional module contains one or more deconvolution, activation function, and random inactivation layers; the last deconvolutional module adopts a softmax two-class activation function, and the output of the deconvolutional network is a 0-1 classification matrix of a single shot record; the 0-1 classification matrix of a single shot record refers to Each sampling point recorded by a single shot is classified as 0 or 1 by a deconvolution network; wherein 0 represents before the first arrival, and 1 represents after the first arrival; the distribution area of ​​the first arrival "before and after arrival" refers to the first arrival area before arrival composed of the set of sampling points classified as 0, and the first arrival area after arrival composed of the set of sampling points classified as 1 in the 0-1 classification matrix output by the deconvolution network; the convolution network and the deconvolution network are trained in the following manner: initializing the convolution kernel parameters of the convolution network and the deconvolution kernel parameters of the deconvolution network, inputting a normalized seismic signal attribute matrix, or multiple normalized seismic signal attribute matrices with the same rows and columns, into the convolution network and the deconvolution network to output a 0-1 classification matrix; quantifying the difference between the predicted first arrival "before and after arrival" distribution area and the manually marked distribution area by a cross entropy loss function, and updating the convolution kernel parameters of the convolution network and the deconvolution kernel parameters of the deconvolution network until the difference value calculated by the cross entropy loss function converges to a smaller value.

2. The earthquake first arrival picking method based on convolutional neural network according to claim 1, characterized in that: The first arrival time is: first, calculate the first arrival sampling point position of each column according to the 0-1 classification matrix output by the deconvolution network; then obtain the sampling time interval in the seismic record file; the multiplication of the two is the first arrival time.

3. The earthquake first arrival picking method based on convolutional neural network according to claim 2 is characterized in that: Assume that the 0-1 classification matrix C has m rows and n columns, which means that the single shot record has n channels, each with m sampling points; then the first arrival sampling point position refers to: the position (i, j) in the column vector, which is the first sampling point position in the column vector that is determined to be classified as 1, and is defined as follows: ； Or it is the starting sampling point position of multiple consecutive sampling positions in the column vector that are determined to be classified as 1, defined as follows: ; Where N means that N consecutive sampling points all belong to category 1; assuming that the first arrival is at sampling point (i, j), k represents the number of the sampling point before the i-th sampling point in channel j; w represents the number of the N consecutive sampling points from sampling point (k, j) to sampling point (k+N-1, j).

Images