阅读说明：本技术 一种机器学习框架下高维特征约束的初至拾取方法及系统 (First arrival picking method and system for high-dimensional feature constraint under machine learning framework ) 是由 张金淼 罗飞 孙文博 朱振宇 翁斌 王小六 郝振江 王艳冬 糜芳 杨俊� 于 2019-10-11 设计创作，主要内容包括：本发明涉及一种机器学习框架下高维特征约束的初至拾取方法及系统,其特征在于,包括以下步骤：1)获取地震数据,并对地震数据进行预处理,得到地震数据的若干地震特征；2)将得到的若干地震特征作为样本输入至预先构建的改进度量方式的模糊聚类算法框架中,对地震数据的地震特征进行筛选,使得地震数据中提取的若干地震特征自动分为初至走时和非初至走时两类,完成地震初至走时的拾取,本发明可以广泛应用于地震勘探领域中。(The invention relates to a first arrival picking method and a system of high-dimensional feature constraint under a machine learning framework, which are characterized by comprising the following steps of: 1) acquiring seismic data, and preprocessing the seismic data to obtain a plurality of seismic characteristics of the seismic data; 2) the obtained seismic characteristics are input into a pre-constructed fuzzy clustering algorithm frame with an improved measurement mode as samples, the seismic characteristics of the seismic data are screened, the seismic characteristics extracted from the seismic data are automatically divided into first arrival travel time and non-first arrival travel time, and the first arrival travel time of the earthquake is picked up.)

1. A first arrival picking method of high-dimensional feature constraint under a machine learning framework is characterized by comprising the following steps:

1) acquiring seismic data, and preprocessing the seismic data to obtain a plurality of seismic characteristics of the seismic data;

2) and inputting the obtained seismic characteristics serving as samples into a pre-constructed fuzzy clustering algorithm frame with an improved measurement mode, and screening the seismic characteristics of the seismic data, so that the seismic characteristics extracted from the seismic data are automatically divided into first arrival travel time and non-first arrival travel time, and the pickup of the first arrival travel time of the earthquake is completed.

2. The machine learning framework high-dimensional feature constraint first arrival picking method according to claim 1, wherein the seismic features comprise energy features, instantaneous travel time features and/or image segmentation based high-dimensional travel time features.

3. The machine learning framework high-dimensional feature constraint first arrival picking method according to claim 2, wherein the energy features are as follows:

wherein, e (i) is energy characteristic, d is seismic data, i is a sampling point, w is time window size, and e (i) is energy of ith point in the seismic data.

4. The machine learning framework high-dimensional feature constraint first arrival picking method according to claim 2, wherein the instantaneous travel time features are as follows:

where ω is the angular frequency, Δ ω is the frequency sampling interval, ω_maxIs the maximum frequency, ω_minFor minimum frequency, T (T, ω) is a time-frequency characterization of instantaneous travel time, and:

wherein, H (t, omega) is the time-frequency transformation of the seismic data H (t).

5. The machine learning framework-based first arrival picking method for high-dimensional feature constraint according to claim 4, wherein the specific acquisition process of the high-dimensional travel time feature based on image segmentation is as follows:

① the seismic data is converted to a gray scale map by the following equation:

D₂(x,t)＝|D(x,t)|

wherein D is the two-dimensional representation of the seismic data, x is the number of seismic channels, and t is the longitudinal sampling time of the seismic data;

② applying Gaussian filter to gray image D₂(x, t) performing Gaussian smoothing;

③ obtaining the magnitude and direction of each point gradient in the Gaussian smoothing result based on the first order difference to obtain the edge strength and normal vector of each point in the gray level image;

④ determining edge points of the gray-scale map according to the edge intensity and normal vector of each point in the gray-scale map by using non-maximum values;

⑤, connecting each edge point by adopting a double-threshold algorithm to obtain the high-dimensional characteristics based on image segmentation during the first arrival.

6. The first arrival picking method of high dimensional feature constraint under machine learning framework as claimed in claim 5, wherein the construction process of the fuzzy clustering algorithm framework of the improved metric method is as follows:

a) determining a characteristic measurement mode based on wavelet phase;

b) and constructing a fuzzy clustering algorithm framework for improving the measurement mode according to the determined characteristic measurement mode based on the wavelet phase.

7. The method as claimed in claim 6, wherein the wavelet phase-based feature metric in step a) is as follows:

L₂(x_j,c_i)＝max_pcos(x_j,W(c_i,p))

wherein L is₂(x_j,c_i) Is a sample x_jAnd class center c_iThe phase change in the local region, W is a window function and p is the magnitude of the window function.

8. The first arrival picking method of high dimensional feature constraint under machine learning framework as claimed in claim 7, wherein the specific process of step b) is:

I) initializing membership degrees to enable the membership degrees to meet constraint conditions;

II) calculating the class center c of the seismic feature_i：

Wherein u is_ijIs degree of membership, x_jA certain sample, c_iIs the class center of a certain class i, n is the number of samples, and m is a factor of membership degree;

III) calculating the wavelet phase L of the seismic features by adopting a wavelet phase-based feature measurement mode₂(x_j,c_i) If the wavelet phase L of the seismic feature₂(x_j,c_i) Stopping iteration if the distance between the first arrival travel time and the second arrival travel time is less than a preset threshold value, so that the seismic features representing the first arrival travel time and the seismic features representing the non-first arrival travel time in the seismic data are separated, and classification of the seismic features is completed; if the wavelet phase L of the seismic feature₂(x_j,c_i) If the value is not less than the preset threshold value, entering the step IV);

IV) calculating the new degree of membership according to the following formula and entering step II):

wherein, c_kClass center of class k.

9. The first arrival picking method of high dimensional feature constraint under machine learning framework as claimed in claim 8, wherein the constraint conditions in step I) are:

10. a first arrival picking system of high-dimensional feature constraint under a machine learning framework is characterized by comprising:

the earthquake feature acquisition module is used for acquiring earthquake data and preprocessing the earthquake data to obtain a plurality of earthquake features of the earthquake data;

and the seismic characteristic screening module is used for inputting the obtained seismic characteristics serving as samples into a pre-constructed fuzzy clustering algorithm frame of an improved measurement mode, screening the seismic characteristics of the seismic data, automatically dividing the seismic characteristics extracted from the seismic data into a first arrival travel time class and a non-first arrival travel time class, and finishing the pickup of the first arrival travel time of the seismic.

Technical Field

The invention relates to a first arrival picking method and a system for high-dimensional feature constraint under a machine learning framework, belonging to the field of seismic exploration.

Background

The travel time pickup is an important step of seismic data processing, plays an important role in tomography, static correction, velocity analysis, AVO analysis and geological interpretation, and many geophysicists propose algorithms for the travel time pickup, and the methods have respective advantages, disadvantages and application ranges. Therefore, the research of the travel time automatic picking method has practical significance, and the stability and the accuracy of the algorithm have important value in economy.

Conventional time-lapse picking methods can be broadly classified into a sliding time window method and a phase drying method (Molyneux and Schmitt, 1999). In the sliding time window method, the attributes of the seismic signal sequence are calculated in continuous or overlapped moving windows, the coherent method relies on comparing single or multiple waveforms using some similarity measurement methods, and in recent years, machine learning algorithms have become more and more widely used in the field of geophysical exploration due to the rapid development of artificial intelligence. In view of the fact that most of the traditional travel time picking methods are processing ideas based on single or double channels, with the increasing of the seismic data volume, single-channel seismic event extraction is discussed in an isolated mode, the transverse spatial continuity of the data is ignored, and therefore the characteristic information of the data is not fully utilized. Therefore, it is necessary to improve the accuracy of travel time picking in consideration of how to introduce high-dimensional seismic data feature constraints under a machine learning framework and enhance the automation capability of travel time picking.

Disclosure of Invention

In view of the above problems, an object of the present invention is to provide a first arrival picking method and system with high dimensional feature constraints under a machine learning framework, which can enhance the automation capability of travel time picking and improve the accuracy of travel time picking.

In order to achieve the purpose, the invention adopts the following technical scheme: a first arrival picking method of high-dimensional feature constraint under a machine learning framework is characterized by comprising the following steps: 1) acquiring seismic data, and preprocessing the seismic data to obtain a plurality of seismic characteristics of the seismic data; 2) and inputting the obtained seismic characteristics serving as samples into a pre-constructed fuzzy clustering algorithm frame with an improved measurement mode, and screening the seismic characteristics of the seismic data, so that the seismic characteristics extracted from the seismic data are automatically divided into first arrival travel time and non-first arrival travel time, and the pickup of the first arrival travel time of the earthquake is completed.

Further, the seismic features include energy features, instantaneous travel time features, and/or high-dimensional travel time features based on image segmentation.

Further, the energy characteristic is:

wherein, e (i) is energy characteristic, d is seismic data, i is a sampling point, w is time window size, and e (i) is energy of ith point in the seismic data.

Further, the instantaneous travel time characteristic is as follows:

where ω is the angular frequency, Δ ω is the frequency sampling interval, ω_maxIs the maximum frequency, ω_minFor minimum frequency, T (T, ω) is a time-frequency characterization of instantaneous travel time, and:

wherein, H (t, omega) is the time-frequency transformation of the seismic data H (t).

Further, the specific acquisition process of the high-dimensional travel time characteristic based on image segmentation is that ① converts the seismic data into a gray-scale map through the following formula:

D₂(x,t)＝|D(x,t)|

wherein D is the two-dimensional representation of the seismic data, x is the number of seismic channels, and t is the longitudinal sampling time of the seismic data;

② applying Gaussian filter to gray image D₂(x, t) performing Gaussian smoothing processing, ③ obtaining the gradient amplitude and direction of each point in the Gaussian smoothing result based on first-order difference to obtain the edge intensity and normal vector of each point in the gray-scale image, ④ determining the edge points of the gray-scale image according to the edge intensity and normal vector of each point in the gray-scale image by adopting non-maximum values, ⑤ connecting the edge points by adopting a double-threshold algorithm to obtain the high-dimensional characteristic based on image segmentation during first-arrival walking.

Further, the construction process of the fuzzy clustering algorithm framework of the improved metric method comprises the following steps: a) determining a characteristic measurement mode based on wavelet phase; b) and constructing a fuzzy clustering algorithm framework for improving the measurement mode according to the determined characteristic measurement mode based on the wavelet phase.

Further, the wavelet phase-based feature measurement method in the step a) is as follows:

L₂(x_j,c_i)＝max_pcos(x_j,W(c_i,p))

wherein L is₂(x_j,c_i) Is a sample x_jAnd class center c_iThe phase change in the local region, W is a window function and p is the magnitude of the window function.

Further, the specific process of the step b) is as follows: I) initializing membership degrees to enable the membership degrees to meet constraint conditions; II) calculating the class center c of the seismic feature_i：

Wherein u is_ijIs degree of membership, x_jA certain sample, c_iIs the class center of a certain class i, n is the number of samples, and m is a factor of membership degree; III) calculating the wavelet phase L of the seismic features by adopting a wavelet phase-based feature measurement mode₂(x_j,c_i) If the wavelet phase L of the seismic feature₂(x_j,c_i) Stopping iteration if the distance between the first arrival travel time and the second arrival travel time is less than a preset threshold value, so that the seismic features representing the first arrival travel time and the seismic features representing the non-first arrival travel time in the seismic data are separated, and classification of the seismic features is completed; if the wavelet phase L of the seismic feature₂(x_j,c_i) If the value is not less than the preset threshold value, entering the step IV); IV) calculating the new degree of membership according to the following formula and entering step II):

wherein, c_kClass center of class k.

Further, the constraint conditions in step I) are:

a first arrival picking system of high-dimensional feature constraint under a machine learning framework is characterized by comprising: the earthquake feature acquisition module is used for acquiring earthquake data and preprocessing the earthquake data to obtain a plurality of earthquake features of the earthquake data; and the seismic characteristic screening module is used for inputting the obtained seismic characteristics serving as samples into a pre-constructed fuzzy clustering algorithm frame of an improved measurement mode, screening the seismic characteristics of the seismic data, automatically dividing the seismic characteristics extracted from the seismic data into a first arrival travel time class and a non-first arrival travel time class, and finishing the pickup of the first arrival travel time of the seismic.

Due to the adoption of the technical scheme, the invention has the following advantages: 1. according to the invention, on the basis of traditional one-dimensional seismic feature picking-up time, high-dimensional seismic data features (attributes) are introduced, so that the transverse continuity of seismic data is considered, the picked first-arrival travel time has more physical significance, and the precision of automatic picking-up first-arrival travel time is improved. 2. The invention adopts a characteristic measurement mode based on wavelet phase to replace the characteristic measurement mode based on Euclidean distance in the traditional fuzzy clustering algorithm, so that the measurement mode in the machine learning algorithm has more physical significance, and the precision of the fuzzy clustering analysis algorithm is improved. 3. The method organically combines several screened seismic characteristics sensitive to first arrival travel time with a machine learning algorithm framework, improves the information utilization rate of seismic data during processing and interpretation, is beneficial to automatic extraction of seismic travel time, has important theoretical and application values, and has certain promotion effect on development of oil gas and mineral resource exploration.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a comparison diagram of several seismic features illustrated in the present invention, wherein FIG. 2(a) is a certain synthesized original seismic data, FIG. 2(b) is a seismic data after the seismic data of FIG. 2(a) is subjected to noise processing, FIG. 2(c) is a diagram of energy features of the seismic data, FIG. 2(d) is a diagram of instantaneous travel-time features of the seismic data, and FIG. 2(f) is a diagram of high-dimensional travel-time features of the seismic data based on image segmentation;

fig. 3 is a schematic diagram of the first arrival picking result of the seismic data obtained by the method of the present invention in the embodiment of the present invention, wherein the black line represents the picking result.

Detailed Description

The present invention is described in detail below with reference to the attached drawings. It is to be understood, however, that the drawings are provided solely for the purposes of promoting an understanding of the invention and that they are not to be construed as limiting the invention.

As shown in fig. 1, the first arrival picking method of high-dimensional feature constraint under a machine learning framework provided by the present invention includes the following steps:

1) the method comprises the following steps of obtaining seismic data, preprocessing the seismic data to obtain a plurality of seismic characteristics of the seismic data, and specifically comprises the following steps:

as shown in fig. 2(a) and (b), in order to obtain seismic data and to apply noise processing to the seismic data, for a machine learning framework, a number of seismic features sensitive to first arrival travel time information need to be input, and the features should have universality. For the first arrival travel time of seismic data, the energy difference before and after the first arrival is the largest. Therefore, the invention screens the following seismic features for matching and is used for fuzzy clustering analysis.

1.1) energy characteristics

For the acquired seismic data, the first arrival travel time position is generally defined as the takeoff or peak position of the seismic wavelet, and therefore, as shown in fig. 2(c), the energy e (i) change of the seismic data is a good first arrival travel time scaling feature:

wherein d is seismic data, i is a certain sampling point, and w is the size of a time window. The meaning of equation (1) is that the energy e (i) at the ith point in the seismic data is equal to the sum of the squares of the local range (i-w, i + w) values at the ith point in the seismic data. The energy characteristic obtained based on the formula (1) can amplify the energy at the first arrival travel time, so that the first arrival travel time characteristic is more obvious and is beneficial to identification.

1.2) instantaneous travel time characteristics

The traditional extraction of the instantaneous travel time characteristics is based on Fourier transform, the Fourier transform is only suitable for stable data, and the seismic data are generally non-stable data, so that the concept of time-frequency transform is introduced, the travel time characteristics are local relative to time, meanwhile, the arrival time of different signals is mutually isolated, and the extracted travel time characteristics are more accurate. Based on this, as shown in fig. 2(d), the instantaneous travel time characteristic t (t) can be redefined as:

h (t, omega) is time-frequency transformation of the seismic data H (t), Im { } is an imaginary part, omega is angular frequency, delta omega is a frequency sampling interval, and omega is frequency_maxIs the maximum frequency, ω_minIs the minimum frequency, T (T) is the instantaneous travel time characteristic of the seismic data, T (T, omega) is the time-frequency representation of the instantaneous travel time, and:

the calculation of the instantaneous travel time of the band-limited signal according to equations (2) and (3) can be divided into two steps of non-interfering ① mapping a certain seismic data to the time-frequency domain by equation (2), ② remapping the time-frequency domain T (T, ω) back to the time domain T (T) by equation (3) weighted along the frequency direction.

1.3) high-dimensional travel time characteristics based on image segmentation

The image segmentation method is widely applied to signal analysis and image processing, and the energy of the seismic data at the first arrival travel time generally has a jump to form a more obvious edge. As the name suggests, the edge is a high-dimensional concept, so that a rough edge formed in a high-dimensional space by the first-arrival travel time can be sketched out based on a Canny edge detector (1986), and the high-dimensional travel time characteristic can be obtained.

1.3.1) converting seismic data to a gray scale map by the following equation (4):

D₂(x,t)＝|D(x,t)| (4)

wherein D is a two-dimensional representation of the seismic data, x is the number of seismic traces, and t is the longitudinal sampling time of the seismic data.

1.3.2) use of Gaussian filters for the grayscale map D₂(x, t) performing Gaussian smoothing.

1.3.3) obtaining the amplitude and direction of the gradient (i.e. the first-order partial derivative of each point) of each point in the Gaussian smoothing result based on the first-order difference to obtain the edge strength and normal vector of each point in the gray-scale map.

1.3.4) adopting non-maximum value, and determining the edge point of the gray-scale map according to the edge intensity and normal vector of each point in the gray-scale map.

1.3.5) adopting a double-threshold algorithm (Otsu method), checking and connecting each edge point to obtain high-dimensional characteristics based on image segmentation during first arrival walking.

As shown in fig. 2(f), a schematic diagram of the high-dimensional features of the two-dimensional seismic data obtained by the above method is just compared with the first two one-dimensional seismic features, so that the high-dimensional features are more sensitive to first arrival travel time and the obtained result is more accurate.

The three characteristics which are sensitive to the first arrival travel time of the earthquake are introduced, but the three characteristics are not limited to be used as the input of the fuzzy clustering algorithm, and two characteristics can be selected for the earthquake data with obvious characteristics. For complex seismic data, more seismic features related to first-arrival travel time can be searched according to the above thought, but the introduction of high-dimensional seismic features can more reasonably pick up the seismic first-arrival travel time.

2) And constructing a fuzzy clustering algorithm framework for improving the measurement mode (distance).

The neural network method belongs to supervised learning, needs a large amount of sample input, and the cluster analysis data is unsupervised learning and can be automatically divided into two types (seismic events and non-seismic events) according to the characteristics of the data. Therefore, the Fuzzy clustering algorithm is selected as a framework, and the principle of the Fuzzy clustering algorithm (FCM, Fuzzy C-means) is as follows:

assuming that the data set is X, if the data set is divided into C classes, corresponding to C class centers, the membership degree of each sample j belonging to a certain class i is u_ijAn objective function J of an FCM and its constraints are defined as follows:

wherein the degree of membership u_ijIs a measure of how similar a sample is to different results; x is the number of_jIs a certain sameThen, the process is carried out; c. C_iClass center for a class i; n is the number of samples. The above formula (5) is formed by multiplying the membership degree of the corresponding sample by the distance (euclidean distance) from the sample to the center of each class, and m is a factor of the membership degree (belonging to the degree of slowness of a certain sample). Equation (6) is a constraint that the sum of the membership degrees of all classes to which a sample belongs is 1.

Using the lagrange multiplier method to solve equation (5) above with constraints, equation (5) above can be written as:

objective function J pair membership u_ijCalculating a partial derivative, and simplifying the partial derivative into:

eliminating lambda according to the formula (6) to obtain the membership degree u_ij：

Wherein, c_kClass center for class k; the numerator of the above equation (9) represents the class center distance of a point with respect to a class, the denominator represents the sum of the class center distances of the point with respect to all classes, and the division of the two represents the specific gravity of the sum of the distance from the point to the class center at the point to all class centers. When the molecules within the sum are smaller, closer to the class, u_ijThe larger the more data belongs to the class.

Similarly, the objective function J is centered on the class center c_iCalculating a partial derivative, and simplifying the partial derivative into:

the above formula (10) is a class-centered updating rule, which is essentially a weighted average of all points, and the weighting coefficient

Multiplied by x_jI.e. the contribution value of the point to class i.

The general steps of the conventional FCM algorithm are:

i initializes the degree of membership u so that it satisfies the constraint of equation (6).

II calculate class center c using equation (10) above_i。

III, calculating a target function J, and stopping iteration if the target function J is smaller than a preset threshold value; and if the target function J is not less than the preset threshold value, entering the step IV.

IV the new degree of membership u is calculated according to equation (9) and step II is entered.

Based on the principle, the specific process of constructing the fuzzy clustering algorithm framework for improving the measurement mode comprises the following steps:

2.1) determining a characteristic measurement mode based on the wavelet phase.

As is well known, a large part of machine learning algorithms introduce the concept of distance to discriminate the differences of different classes, such as cluster analysis. From the above equation (5), the measurement of the conventional FCM algorithm is the distance L from the sample to the center of each class₁(x_j,c_i) (Euclidean distance):

L₁(x_j,c_i)＝||x_j-c_i||²(11)

on the basis, the invention provides a new measurement mode based on wavelet phase characteristics, which is shown in the following formula (12):

L₂(x_j,c_i)＝max_pcos(x_j,W(c_i,p)) (12)

wherein L is₂(x_j,c_i) Is a sample x_jAnd class center c_iThe phase change in the local region, W is a window function and p is the magnitude of the window function. The above equation (12) is a new wavelet phase characteristic-based metric, and is significantly different from the equation (11), i.e. the conventional euclidean distance metric form. Article thereofThe rational meaning is that formula (12) not only ensures that the sample is the shortest from a certain class, but also ensures that sample x_jAnd class center c_iThe phase of the wavelets in the local region remains consistent.

2.2) constructing a fuzzy clustering algorithm framework of an improved measurement mode according to the determined characteristic measurement mode based on the wavelet phase, namely formula (12).

The seismic data all include wavelet phase characteristics, and the wavelet phases of the seismic event and the non-seismic event are obviously different. Therefore, the invention replaces the objective function J (i.e. formula (5)) in step III in the traditional FCM algorithm flow with formula (12), and constitutes a fuzzy clustering algorithm framework for improving the metric method. Only when the phase difference of the sub-waves is not large when the sample is close to the class center, the residual error can be minimized, the seismic data can be automatically divided into two classes (seismic events and non-seismic events), and meanwhile, the classification is ensured to have physical significance and higher precision, namely, the fuzzy clustering algorithm framework of the improved measurement mode comprises the following contents:

2.2.1) initializing the membership degree u so that the membership degree u meets the constraint condition of the formula (6).

2.2.2) calculating the class center c of the seismic signature using equation (10) above_i。

2.2.3) calculating the wavelet phase L of the seismic features by adopting a wavelet phase-based feature measurement mode₂(x_j,c_i) If the wavelet phase L of the seismic feature₂(x_j,c_i) Stopping iteration if the distance between the first arrival travel time and the second arrival travel time is less than a preset threshold value, so that the seismic features representing the first arrival travel time and the seismic features representing the non-first arrival travel time in the seismic data are separated, and classification of the seismic features is completed; if the wavelet phase L of the seismic feature₂(x_j,c_i) If not, step 2.2.4) is performed, where the threshold may be set according to the actual situation, which is not described herein.

2.2.4) calculating a new membership u according to the above formula (9) and then proceeding to step 2.2.2).

3) Taking a plurality of obtained seismic features as a sample x_jInput to the constructed fuzzy clustering algorithm frameworkIn the method, the seismic characteristics of the seismic data are screened, so that a plurality of seismic characteristics extracted from the seismic data are automatically divided into first arrival travel time and non-first arrival travel time, and the first arrival travel time of the earthquake is picked up.

X in the above equation (5) in combination with seismic characteristics_jAnd c_jIt should be a high dimensional function, and a certain component of the function represents a feature, and its dimension is the number of the features. According to several earthquake characteristics given in the step 1), the improved fuzzy clustering algorithm of the formula (12) is adopted, so that the earthquake data can be fully automatically divided into first-arrival travel time and non-first-arrival travel time, and the purpose of automatically picking the earthquake first-arrival travel time is achieved.

As shown in fig. 3, which is a schematic diagram of the first arrival picking result obtained by applying the method of the present invention to a certain actual seismic data in BZ29 region, it can be seen that the method of the present invention can successfully pick the first arrival travel time information of 450-550 channels (absorption attenuation, weak energy region) and the first arrival travel time information of 600 channels (low signal-to-noise ratio region). The method not only considers the transverse continuity of the seismic data, but also considers the consistency of the wavelets, and compared with the traditional first arrival picking method, the method has higher picking precision and more physical significance.

Based on the first arrival picking method of the high-dimensional feature constraint under the machine learning framework, the invention also provides a first arrival picking system of the high-dimensional feature constraint under the machine learning framework, which comprises the following steps:

the earthquake feature acquisition module is used for acquiring earthquake data and preprocessing the earthquake data to obtain a plurality of earthquake features of the earthquake data; and the seismic characteristic screening module is used for inputting the obtained seismic characteristics serving as samples into a pre-constructed fuzzy clustering algorithm frame of an improved measurement mode, screening the seismic characteristics of the seismic data, automatically dividing the seismic characteristics extracted from the seismic data into a first arrival travel time class and a non-first arrival travel time class, and finishing the pickup of the first arrival travel time of the seismic.

The above embodiments are only used for illustrating the present invention, and the structure, connection mode, manufacturing process, etc. of the components may be changed, and all equivalent changes and modifications performed on the basis of the technical solution of the present invention should not be excluded from the protection scope of the present invention.