CN-121995527-A - Method and device for simultaneously realizing stratum comparison, horizon interpretation and speed modeling

CN121995527ACN 121995527 ACN121995527 ACN 121995527ACN-121995527-A

Abstract

The application provides a method and a device for simultaneously realizing stratum contrast, horizon interpretation and speed modeling, and relates to the technical field of oil and gas exploration and artificial intelligence intersection. The method comprises the steps of obtaining well tracks, logging data and seismic data in a work area to be modeled, extracting corresponding parawell seismic channels from the seismic data along the well tracks through well shock calibration, aligning the logging data and the parawell seismic channels to a unified depth domain or time domain to form a multichannel input sequence, inputting the multichannel input sequence into a pre-trained neural network model to obtain probability distribution of each stratum category, and inputting the probability distribution of each stratum category into a conditional random field to perform global optimization to generate stratum modeling results conforming to geological deposition rules. According to the method, the multi-channel input sequence is constructed by fusing logging data and the side-hole seismic channels, so that the limitation of comprehensive interpretation of single data source stratum is overcome, and a multi-task comprehensive interpretation result of stratum comparison, horizon interpretation and speed modeling with geological rationality is generated.

Inventors

LI QIXIN
Yuan sanyi
WANG BO
LIU LANG
LI KUNHAN

Assignees

中国石油大学(北京)

Dates

Publication Date: 20260508
Application Date: 20260313

Claims (10)

1. A method for simultaneously implementing formation contrast, horizon interpretation and velocity modeling, comprising: acquiring well tracks, logging data and seismic data in a work area to be modeled; Extracting a corresponding parawell seismic trace from the seismic data along the well trace through well seismic calibration; Aligning the logging data with the parawell seismic traces to a uniform depth domain or time domain to form a multichannel input sequence; Inputting the multichannel input sequence into a pre-training neural network model to obtain probability distribution of each stratum category, wherein the pre-training neural network model comprises an encoder, a decoder and an output layer, the encoder comprises a plurality of downsampling stages, a converter is embedded in a deep layer stage to model long-range dependency of the stratum sequence, the decoder adopts an upsampling structure and fuses high-resolution characteristics output by the encoder through jump connection, and the output layer outputs the probability distribution of each stratum category through one-dimensional convolution and an activation function; Inputting probability distribution of each stratum category into a conditional random field CRF for global optimization, and generating stratum modeling results conforming to geological deposition rules; Separately inputting logging data of a depth domain to the pre-training neural network model for a target well position in the work area to be modeled, and generating a first stratum sequence label for stratum comparison; Independently inputting a time domain parawell seismic channel to the pre-training neural network model for the target well position to generate a second stratum sequence label for horizon interpretation; Matching the interfaces with the same stratum category in the first stratum sequence tag and the second stratum sequence tag along the longitudinal direction of the well track, determining the corresponding depth position and time position according to the interfaces, and constructing the time-depth relation of the target well position; And according to the time-depth relation, calculating the layer speed of each stratum through the depth difference between adjacent stratum interfaces and the double-way travel time difference, and inverting to obtain a layer speed model.
2. The method of claim 1, wherein the training process of the pre-trained neural network model comprises: Acquiring historical work area Extracting a corresponding second well-side seismic trace from the second seismic data along the second well trace through well-shock calibration, and aligning the second well trace, the second well-logging data, the second seismic data and the neural network model to be trained to a uniform depth domain or a uniform time domain to form a multichannel training input sequence; And inputting the multichannel training input sequence into the neural network model to be trained for training, and obtaining the pre-training neural network model.
3. The method according to claim 2, wherein said inputting the multi-channel training input sequence into the neural network model to be trained to obtain the pre-trained neural network model comprises: Setting part of input channels in the multi-channel training input sequence to zero according to preset probability to obtain a simulation training sequence so as to simulate a scene of missing logging data or seismic data; and inputting the simulated training sequence into the neural network model to be trained for training, and obtaining the pre-training neural network model.
4. A method according to claim 2 or 3, further comprising: Adopting a weighted sum of cross entropy loss and dess price loss as a mixed loss function in the training process, wherein the weight coefficient of the cross entropy loss is a first weight coefficient, the weight coefficient of the price loss is a second weight coefficient, and the sum of the first weight coefficient and the second weight coefficient is 1; And updating parameters of the neural network model to be trained by an adaptive moment estimation optimizer Adam at a preset initial learning rate to obtain the pre-training neural network model.
5. The method of claim 1, wherein the inputting the probability distribution of each formation class into the conditional random field CRF for global optimization generates a formation modeling result conforming to a geological deposition law, comprising: Constructing a linear chain CRF, wherein node potential energy is determined by probability distribution of each stratum category, and side potential energy models transition probabilities among stratum categories based on geological deposition constraints; In the transition probabilities, the transition from the formation category of the older geologic age to the formation category of the older geologic age is prohibited, and the transition between the same formation categories or the transition from the formation category of the older geologic age to the formation category of the newer geologic age is allowed; And solving a global optimal stratum tag sequence of the linear chain CRF through a Viterbi algorithm, and taking the global optimal stratum tag sequence as the stratum modeling result.
6. An apparatus for simultaneously implementing formation contrast, horizon interpretation and velocity modeling, comprising: The acquisition module is used for acquiring well tracks, logging data and seismic data in the work area to be modeled; the extraction module is used for extracting a corresponding parawell seismic channel along the well track from the seismic data through well seismic calibration; the alignment module is used for aligning the logging data and the parawell seismic channels to a unified depth domain or time domain to form a multichannel input sequence; an input module for inputting the multi-channel input sequence into a pre-training neural network model to obtain probability distribution of each stratum category, wherein the pre-training neural network model comprises an encoder, a decoder and an output layer, the encoder comprises a plurality of downsampling stages, a converter is embedded in a deep stage to model long-range dependency of stratum sequence, the solution The encoder adopts an up-sampling structure and fuses high-resolution characteristics output by the encoder through jump connection, and the output layer outputs probability distribution of each stratum category through one-dimensional convolution and an activation function; The generation module is used for inputting probability distribution of each stratum category into a Conditional Random Field (CRF) to perform global optimization, and generating stratum modeling results conforming to a geological deposition rule; an extension module is applied and used for independently inputting logging data of a depth domain to the pre-training neural network model for a target well position in the work area to be molded to generate a first stratum sequence label for stratum comparison; The application extension module is further used for independently inputting a side-of-well seismic channel of a time domain to the pre-training neural network model for the target well position to generate a second stratum sequence label for horizon interpretation; The application extension module is further used for matching the interfaces with the same stratum category in the first stratum sequence tag and the second stratum sequence tag along the longitudinal direction of the well track, determining the corresponding depth position and the corresponding time position according to the interfaces, and constructing the time-depth relation of the target well position; The application extension module is further used for calculating the layer speed of each stratum according to the time-depth relation through the depth difference between adjacent stratum interfaces and the double-way travel time difference, and inverting to obtain a layer speed model.
7. The apparatus of claim 6, further comprising a pre-training module; the pre-training module is used for acquiring a second well track, second well logging data, second seismic data and a neural network model to be trained in the historical work area; The pre-training module is further configured to extract a corresponding second parawell seismic trace from the second seismic data along the second well trace through well seismic calibration, and align the second well logging data and the second parawell seismic trace to a uniform depth domain or a uniform time domain, so as to form a multi-channel training input sequence; The pre-training module is further configured to input the multi-channel training input sequence to the neural network model to be trained to obtain the pre-training neural network model.
8. An electronic device is characterized by comprising a memory and a processor; The memory stores computer-executable instructions; The processor executing computer-executable instructions stored in the memory, causing the processor to perform the method of any one of claims 1 to 5.
9. A computer readable storage medium having stored therein computer executable instructions which when executed by a processor are adapted to carry out the method of any one of claims 1 to 5.
10. A computer program product comprising a computer program which, when executed by a processor, implements the method of any one of claims 1 to 5.

Description

Method and device for simultaneously realizing stratum comparison, horizon interpretation and speed modeling Technical Field The application relates to the technical field of intersection of oil and gas exploration and artificial intelligence, in particular to a method and a device for simultaneously realizing stratum contrast, horizon interpretation and speed modeling. Background In oil and gas exploration, formation comparison, horizon interpretation and velocity modeling are the core links for developing regional formation interpretation and analysis. The three tasks are regarded as independent processes in the prior art, and the technical problems are that firstly, for stratum comparison tasks, the traditional method relies on geology specialists to manually judge lithology combination and deposit rotation characteristics of a well logging curve, subjective difference is large, efficiency is low, large-scale dense well pattern data are difficult to adapt, the traditional machine learning and deep learning method mainly relies on the well logging curve, for example, a traditional U-shaped network (U-Net) and a segmentation network (Segmentation Network and SegNet) focus on the well logging data, and the well logging curve has high vertical resolution but does not have transverse continuity, so that stratum comparison can only be developed around a well shaft and cannot extend in a three-dimensional space. Secondly, for horizon interpretation tasks, the traditional method relies on experts to manually track the seismic event, which is time-consuming, laborious and poor in consistency. The existing deep learning method mostly adopts a Two-Dimensional U-Net (2D U-Net) or Three-Dimensional U-Net (3D U-Net) architecture, training and prediction are only carried out on seismic data, and the method can utilize the transverse continuity of the earthquake but lacks the direct vertical constraint of logging geologic layering, so that horizon interpretation and geologic layering have poor matching performance at well points. Mutual constraint among different layers cannot be cooperatively considered, flying spots are easy to occur in the section of the anti-reflection layer by single-layer explanation, and layer penetration or misconnection is easy to occur in multi-layer explanation. Finally, for the speed modeling task, the traditional method relies on the sound wave time difference and the density logging curve to build the time depth relation through the synthesis record calibration, or generates the full-area speed field through the interpolation method, but a large amount of well drilling in actual production is limited by economic cost or well conditions to lack the sound wave-density curve, so that the traditional method is difficult to implement. The existing deep learning speed modeling method is mostly based on seismic data or independent logging data, and collaborative modeling of stratum contrast and horizon interpretation tasks under a unified frame is not achieved yet. Therefore, the existing method is fractured in data modes, modeling targets and methods, so that geological-geophysical information cannot be fused deeply, and accuracy and efficiency of stratum interpretation and analysis are restricted. Disclosure of Invention The application provides a method and a device for simultaneously realizing stratum contrast, horizon interpretation and speed modeling, which are characterized in that a multichannel input sequence is constructed by fusing logging data and a parawell seismic channel, an encoder and decoder network is adopted, a transducer (transducer) is embedded in a deep layer to capture long-range dependence, and then a conditional random field (Conditional Random Field, CRF) is used for globally optimizing probability distribution, so that a geological deposition rule is forcibly satisfied, a stratum modeling result with physical consistency is generated, and the precision and the efficiency of stratum interpretation and analysis are improved. In a first aspect, the present application provides a method for simultaneously implementing formation comparison, horizon interpretation and velocity modeling, the method comprising: acquiring well tracks, logging data and seismic data in a work area to be modeled; Extracting a corresponding parawell seismic channel from the seismic data along a well track through well seismic calibration; Aligning logging data with a side-of-well seismic channel to a unified depth domain or time domain to form a multichannel input sequence; Inputting a multichannel input sequence into a pre-training neural network model to obtain probability distribution of each stratum category, wherein the pre-training neural network model comprises an encoder, a decoder and an output layer, the encoder comprises a plurality of downsampling stages, a converter is embedded in a deep layer stage to model long-range dependency of the stratum sequence, the decoder adopts an upsamplin