Multi-scale unsupervised seismic velocity inversion method based on autoencoder for observation data

Patent metadata

Field	Value
Publication number	US-11828894-B2
Application number	US-202118031289-A
Country	US
Kind code	B2
Filing date	Dec 14, 2021
Priority date	Nov 19, 2021
Publication date	Nov 28, 2023
Grant date	Nov 28, 2023

Abstract

A multi-scale unsupervised seismic velocity inversion method based on an autoencoder for observation data. Large-scale information is extracted by the autoencoder, which is used for guiding an inversion network to complete the recovery of different-scale features in a velocity model, thereby reducing the non-linearity degree of inversion. A trained encoder part is embedded into the network to complete the extraction of seismic observation data information at the front end, so it can better analyze the information contained in seismic data, the mapping relationship between the data and velocity model is established better, then the inversion method is unsupervised, and location codes are added to the observation data to assist the network in perceiving the layout form of an observation system, which facilitates practical engineering application. Thus a relatively accurate inversion result of the seismic velocity model when no real geological model serves as a network training label can be achieved.

First claim

What is claimed is:

1. A multi-scale unsupervised seismic velocity inversion method based on an autoencoder for observation data, comprising:
constructing a corresponding geological velocity model according to the actual geological conditions, calculating corresponding simulated seismic observation data by numerical simulation, and based on each seismic observation data and the geological velocity model, forming an unsupervised seismic velocity inversion database;
training a plurality of autoencoders by using the simulated seismic observation data, the different autoencoders encoding global key information in the seismic observation data into low-dimensional vectors of different lengths;
adding a location feature information code to each seismic trace of actual observation data, the code being used for determining location information of seismic observation data of each trace;
constructing a convolutional-fully connected network, embedding each trained encoder part of the autoencoder for observation data into a front end of the above-mentioned network structure, so that an inversion network can effectively extract global information of the observation data at an input end of the seismic data, inputting the seismic observation data encoded by the location feature information into the convolutional-fully connected network, and outputting a predicted velocity model corresponding to the seismic observation data;
constructing a forward modeling network of wave equation to transform the predicted velocity model into corresponding predicted observation data;
calculating a residual of the predicted observation data and the simulated seismic observation data; respectively inputting the predicted observation data and the simulated seismic observation data into each trained encoder part of the autoencoder for observation data to respectively obtain encoded low-dimensional vectors and calculate a residual of each low-dimensional vector; calculating a residual of a linear gradient velocity model and the predicted velocity model output by the inversion network;
summing the three residuals according to the proportion changing with the number of training rounds and then forming a multi-scale unsupervised loss function, guiding the network to recover different-scale information in the velocity model at different training stages, and updating parameters of the convolutional-fully connected network by means of gradient pass back of the loss function; and
using the convolutional-fully connected network with the updated parameters to process field observation data to obtain an inversion result.

2. The multi-scale unsupervised seismic velocity inversion method based on the autoencoder for the observation data according to claim 1 , wherein when the corresponding geological velocity models are constructed according to the actual geological conditions, wavefield simulation is performed for each geological velocity model with a fixed location of seismic sources and receivers, as well as observation time, and wavefield data is recorded at the receiver location to obtain actual seismic data corresponding to the geological velocity models.

3. The multi-scale unsupervised seismic velocity inversion method based on the autoencoder for the observation data according to claim 1 , wherein a number of parameters of all the low-dimensional vectors is less than that of the seismic observation data, and the vector with a smaller number of parameters corresponds to larger-scale information in the velocity model.

4. The multi-scale unsupervised seismic velocity inversion method based on the autoencoder for the observation data according to claim 1 , wherein actual observation data is self-encoded by a plurality of regularization autoencoders, an encoder part output of the regularization autoencoder consisting of a vector with a number of parameters lower than that of the seismic observation data, and the vector containing global key information in the seismic observation data and corresponding to large-scale information in the velocity model.

5. The multi-scale unsupervised seismic velocity inversion method based on the autoencoder for the observation data according to claim 1 , wherein a trigonometric function encoded location feature is added to each seismic trace of the actual observation data, the location feature information code being two numerical values solved by a formula composed of a sine-cosine function input through a shot point of the seismic trace and a receiver location and can realize the calibration of any location of seismic sources and receivers.

6. The multi-scale unsupervised seismic velocity inversion method based on the autoencoder for the observation data according to claim 1 , wherein the convolutional-fully connected network comprises a feature encoder, a feature generator and a feature decoder, each encoder part of the autoencoder for observation data forming the feature encoder together with other network structures, and the network being used for establishing mapping of observed seismic observation data to the velocity model.

7. The multi-scale unsupervised seismic velocity inversion method based on the autoencoder for the observation data according to claim 6 , wherein the encoder comprises a global feature encoder and a neighborhood information encoder, the observation data input into the network is respectively input to the above two parts, and outputs of the two parts are input to the feature generator after being spliced; the global feature encoder is each encoder part of the autoencoder for observation data, and the neighborhood information encoder consists of a 3-layer successively cascaded convolutional structure;
or the feature generator comprises 5 fully connected layers, an input of the feature generator being an output of the encoder, and an output of the feature generator being an input of the feature decoder;
or the feature decoder comprises a 6-layer successively cascaded convolutional structure, wherein the 4th-layer convolutional structure is 4 parallel convolutional layers.

8. The multi-scale unsupervised seismic velocity inversion method based on the autoencoder for the observation data according to claim 1 , wherein in the specific process of constructing the forward modeling network of the wave equation to transform the predicted velocity model into corresponding seismic observation data: constructing the forward modeling network of the wave equation based on a deep neural network, and performing forward modeling of a seismic wavefield on a final output of the convolutional-fully connected network so that the seismic observation data corresponding to the predicted velocity model is obtained.

9. The multi-scale unsupervised seismic velocity inversion method based on the autoencoder for the observation data according to claim 1 , wherein the specific process of constructing the forward modeling network of the wave equation based on the deep neural network comprises: in a time-space domain, discretizing a constant density acoustic wave equation, the process of the seismic wavefield propagating with time being based on the iterative process of a forward operator in the discretized equation; and taking seismic wavefield propagation operation at each time step as a layer of deep neural network, taking the seismic velocity model as a trainable parameter of the deep neural network, and taking convolution operation in the wavefield propagation process and operation between corresponding elements of a matrix as the internal operation process of the network to realize construction of the forward modeling network of the wave equation.

10. The multi-scale unsupervised seismic velocity inversion method based on the autoencoder for

Assignees

• Univ Shandong

Inventors

• Liu Bin
• Ren Yuxiao
• Jiang Peng
• Yang Senlin
• Wang Qingyang
• Xu Xinji
• Li Duo

Classifications

• G01V1/303 Primary

  for determining velocity profiles or travel times · CPC title

• G01V2210/6222

  Velocity; travel time · CPC title

• G01V2210/66

  Subsurface modeling · CPC title

• G06F30/27 Primary

  using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model · CPC title

• G06N3/049

  Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs · CPC title