CN-122017958-A - Shale oil and rock phase combined earthquake phase control inversion method based on deep learning

CN122017958ACN 122017958 ACN122017958 ACN 122017958ACN-122017958-A

Abstract

The invention discloses a shale oil and rock facies combined earthquake phase control inversion method based on deep learning, which comprises the steps of preprocessing a pre-stack angle trace set and logging data, and dividing earthquake phases for the earthquake data by adopting a K-Means clustering algorithm; the method comprises the steps of establishing a three-dimensional convolution depth learning model based on a space-time attention mechanism, converting a seismic phase classification result into a time sequence code by introducing a position code theory in natural language processing, introducing the time sequence code into the three-dimensional convolution depth learning model to form a seismic phase control depth learning model, training the seismic phase control depth learning model by adopting a semi-supervised learning method, and obtaining the transverse resolution of an inversion longitudinal and transverse wave speed and density result. The longitudinal and transverse wave speed and the density result transverse resolution of the phased depth learning inversion are superior to those of the inversion result without the phased, and the thickness of the inverted thin layer is far less than one quarter of the wavelength of the seismic wave.

Inventors

LIU HAOJIE
YU WENZHENG
GAI PANPAN
DING KUN
YUAN HAIHAN
DU ZHENG
BAI WAN

Assignees

中国石油化工股份有限公司
中国石油化工股份有限公司胜利油田分公司

Dates

Publication Date: 20260512
Application Date: 20241112

Claims (11)

1. The shale oil and rock phase combination seismic phased inversion method based on deep learning is characterized by comprising the following steps of: Preprocessing the pre-stack angle trace set and the logging data, and dividing the seismic data into seismic phases by adopting a K-Means clustering algorithm; Establishing a three-dimensional convolution depth learning model based on a space-time attention mechanism, introducing a position coding theory in natural language processing to convert a seismic phase classification result into a time sequence code, and introducing the time sequence code into the three-dimensional convolution depth learning model to form a seismic phase control depth learning model; And training the earthquake phase control deep learning model by adopting a semi-supervised learning method to obtain the inversion longitudinal and transverse wave speed and the density result transverse resolution.
2. The deep learning-based shale oil-rock phase combined seismic facies control inversion method of claim 1, wherein the preprocessing of pre-stack angle trace sets and logging data specifically comprises: and (5) carrying out standardization and normalization treatment on the pre-stack angle trace set and the logging data.
3. The shale oil rock phase combined seismic phased inversion method based on deep learning according to claim 2, wherein the normalization and normalization processing of the prestack angle trace set and the logging data specifically comprises: The seismic data are sample data, the longitudinal wave speed, the transverse wave speed and the density provided by the logging data are tag data, and STANDARDSCALER and MinMaxScaler functions are used for carrying out standardization and normalization processing on the pre-stack angle trace set and the logging data so as to eliminate differences among different logging data dimensions and differences among different angle amplitudes of the pre-stack angle trace set; Wherein functions STANDARDSCALER and MinMaxScaler are: Wherein X i is logging data or a prestack angle trace set, X m and X σ are respectively the mean and variance of the logging data and the prestack angle trace set, Y i is normalized data, ymin and Ymax are respectively the minimum and maximum values of the normalized data, and Y i ' is normalized result.
4. The shale oil and rock facies combination seismic facies control inversion method based on deep learning of claim 1, wherein the partitioning of the seismic data by the K-Means clustering algorithm specifically comprises: and inputting the seismic phase data into a deep learning model, fully superposing the prestack angle trace sets to obtain post-stack seismic data, and carrying out seismic phase division on the seismic data of the target layer by using a K-Means clustering algorithm.
5. The shale oil rock phase combination seismic facies control inversion method based on deep learning according to claim 1, wherein the method for converting the position coding theory into the natural language processing into the time sequence coding specifically comprises the following steps: The position coding technology in natural language processing is to code words or characters in a text sequence to represent their position information in a deep learning model, and the position coding is to help the deep learning model understand the relative positions and sequences of different elements in the sequence and capture semantic information in the sequence; Different seismic facies have different sedimentary stratum structures, and a position coding technology is adopted to convert a seismic facies classification result into a spatial position code which is used as a single input of a neural network to play a role of a seismic facies control constraint deep learning model; the theoretical formula of the position coding is as follows: wherein pos represents the category of the seismic facies, i represents the dimension index of the position-coding vector, d model represents the embedding dimension of the model, and PE m represents the position coding of the mth seismic trace.
6. The shale oil-rock phase combined seismic phased inversion method based on deep learning of claim 1, wherein the building of the three-dimensional convolution deep learning model based on a space-time attention mechanism specifically comprises the steps of inputting logging data, seismic data preprocessing and a seismic phase time sequence coding data set into the three-dimensional convolution deep learning model, wherein the three-dimensional convolution deep learning model based on the space-time attention mechanism contains 9 layers in total; The first part comprises three parallel convolution layers with the convolution kernel size and number of the convolution layers being respectively And Wherein the method comprises the steps of The number of the convolution kernels is represented as n, the sizes of the convolution kernels are (x, y, z), and x, y and z are respectively a main line, a cross line and a depth direction of the seismic data, and the sliding step length is 1; extracting spatial features of different spatial scales of seismic data through convolution kernels of different scales, and then carrying out channel splicing on the extracted spatial features to input the spatial features into Further extracting characteristic information among different scales of the seismic data; the spatial attention layer is used for improving the sensitivity of the deep learning model to the spatial characteristics of the seismic data; the second part comprises three parallel convolution layers with the convolution kernel size and number of the convolution layers being respectively And Extracting time sequence features of different scales of seismic data through convolution kernels of different scales, and then carrying out channel splicing on the extracted time sequence features and inputting the channel spliced time sequence features into a computer To further extract timing characteristic information of the seismic data; the time sequence attention layer is used for improving the sensitivity of the deep learning model to time sequence characteristics; Dropout layer, disabling neurons with probability p during training; And the full-connection layer, the input channel number is 9, and the output channel number is 3, and the full-connection layer is used for improving the nonlinear relation fitting capability between the input layer and the output layer and outputting the prediction result of the three-dimensional convolution deep learning fusion network.
7. The shale oil rock phase combined seismic facies control inversion method based on deep learning according to claim 1, wherein the seismic facies control deep learning model specifically comprises: and establishing a deep learning model loss function based on logging and seismic data priori information constraint.
8. The shale oil and rock phase combined seismic phased inversion method based on deep learning of claim 7, wherein the establishing of the deep learning model loss function based on logging and seismic data priori information constraint specifically comprises: Restricting the training back transmission of the deep learning model by three parts of logging data loss, seismic data loss and low-frequency model data loss; the first part of the formula is loss of elastic parameter data in well logging data, a deep learning model is trained by using well logging and seismic data at well points, the input seismic data predicts the elastic parameter data, and loss errors between true values and predicted values of the elastic parameter data at the well points are obtained; Under the constraint of a weight lambda 1 , compensating the middle-high frequency component of the well logging data into the seismic data, breaking through the limit of the effective bandwidth of the seismic data based on the depth learning seismic inversion technology, wherein the inversion result has higher longitudinal resolution, and the larger lambda 1 is, the higher the longitudinal resolution of the prediction result is; the second part of the formula is loss of the seismic data, the convolution model can establish a relation between rock elasticity parameter data and the seismic data, for the unlabeled seismic data, the convolution model is adopted to forward the elasticity parameter data predicted by the deep learning model into a pre-stack seismic trace set, loss errors between the synthesized seismic data and the real seismic data are obtained, the weight lambda 2 controls the constraint degree of the seismic data, and the larger lambda 2 is, the transverse characteristics of a prediction result are more in accordance with the seismic data; The third part of the formula is loss of low-frequency components, loss errors between low-frequency components of a prediction result obtained by using a low-pass filter and low-frequency components obtained by using combined interpolation of earthquake and logging; the objective function of the network is expressed as: M and N are the numbers of the seismic data and the logging data respectively, and N < < M in the training process is used for meeting the condition that training samples in an actual work area are insufficient; in the inversion process of the model data and the actual data, the weight coefficients of the objective function are respectively set as follows λ 1 ＝0.4,λ 2 ＝0.3,λ 3 =0.3。
9. The shale oil rock phase combined seismic facies control inversion method based on deep learning of claim 8, wherein the training of the seismic facies control deep learning model by the semi-supervised learning method specifically comprises: Inputting the pre-stack seismic trace set data, the seismic phase data and the logging data training set into a three-dimensional convolution deep learning model; And training a deep learning model by using the loss function, adopting momentum Adam as an optimization algorithm, and outputting the deep learning model after the loss function tends to be stable.
10. The shale oil rock phase combined seismic facies control inversion method based on deep learning of claim 1, further comprising testing and evaluating the seismic facies control deep learning model.
11. The method for performing the deep learning based shale oil rock phase combined seismic facies control inversion of claim 10, wherein the performing the test and evaluation on the seismic facies control deep learning model specifically comprises: Inputting the test set of pre-stack seismic trace set data, seismic phase data and logging data into the trained seismic phase control deep learning model, and quantitatively evaluating the deep learning model by utilizing a determinable coefficient R 2 ; The determinable coefficient R 2 is expressed as: Wherein y i represents the real label, Representing the mean value of the true value, Representing the predicted value of the network, n being the number of samples.

Description

Shale oil and rock phase combined earthquake phase control inversion method based on deep learning Technical Field The invention relates to the field of petroleum geophysical exploration, in particular to a shale oil and rock phase combined seismic phase control inversion method based on deep learning. Background With the continuous progress of petroleum exploration and development technology, the precision requirement of reservoir prediction is gradually improved, the seismic inversion technology is a key technology for underground reservoir description, dimensionless seismic data are converted into elastic parameter data with physical significance through logging data and other constraint, and even under reasonable assumption conditions, the elastic parameter data can be finally converted into reservoir physical parameter data, and the east camping concave shale oil stratum has the characteristics of thin interlayer, strong non-uniformity and the like, so that the existing certainty and random seismic inversion technology can hardly identify the advantageous shale oil reservoir. The application number is CN202110998575.7 Chinese patent application, which discloses a data and intelligent optimization dual-drive deep learning seismic wave impedance inversion method, and the deep learning network-based seismic inversion method is mentioned, wherein the global optimization method is utilized to perform wave impedance inversion on a part of post-stack seismic data, the data obtained by inversion is utilized to pretrain the deep learning network to learn the mapping relation from the seismic data to the wave impedance, the pretrained network is utilized to guide the global optimization method to invert the wave impedance data of another part of seismic data, the convergence of the wave impedance data to an optimal solution is accelerated, the acquired optimal solution is utilized to perform the optimization on the deep learning network, and the optimized deep learning network is utilized to efficiently realize the inversion of the wave impedance model of the large-scale three-dimensional seismic data. The invention greatly improves the inversion efficiency of the wave impedance model, and realizes the great breakthrough of enabling the application of the global optimization method in the large-scale wave impedance inversion problem under the condition of affordable calculation cost. However, the patent only demonstrates that the deep learning network has strong nonlinear fitting capability, and an important problem of the deep learning network for seismic inversion is that the tags for well shock matching are too few, so that the generalization capability research of the deep learning network should be enhanced. The Chinese patent application No. CN202211206119.5 discloses a sound wave impedance inversion method and system based on well control semi-supervised deep learning, and the method is used for carrying out mean square error operation on predicted sound wave impedance and real sound wave impedance of seismic data with corresponding well logging curves to obtain well logging loss, and carrying out mean square error operation on the predicted sound wave impedance of the seismic data without the corresponding well logging curves to obtain synthetic seismic records through a seismic convolution model and then obtaining the seismic loss. The well logging loss and the earthquake loss are weighted and summed to obtain training loss, and the artificial neural network updates model parameters by optimizing the training loss. Compared with the traditional method, the acoustic impedance inversion method based on deep learning has the advantages of less man-machine interaction and higher intelligent degree, meanwhile, the acoustic impedance inversion method based on deep learning belongs to a semi-supervised learning method, does not need to train an artificial neural network by using a large amount of data, and is reasonable because a geophysical forward model is added, and acoustic impedance inversion results follow the geophysical law. However, when the deep learning network is used for inverting the seismic wave impedance information, the generalization of the deep learning network only adopts a semi-supervised learning method, so that the inversion result of the deep learning network seismic inversion method has low transverse resolution. The Chinese patent application No. CN201310473076.1 discloses a phase control seismic inversion method in geophysical exploration, which has the advantages that the function of seismic phase control in seismic inversion is mentioned, the wave impedance data of a reservoir is obtained by inversion by utilizing the wave impedance curves of seismic data and well logging under the control of seismic phase and geological phase achievements, the longitudinal resolution can be improved, the transverse resolution can be maintained, and the spatial spreading form of the reservoir can b