LSTM Seismic Trace Reconstruction from Random Quantized Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional seismic data reconstruction methods face challenges with noisy and incomplete data due to sparse and irregular sampling, leading to reduced data quality and accuracy, and existing deconvolution techniques suffer from computational inefficiencies and noise amplification.
Innovation Solution
Employing a long short-term memory (LSTM) autoencoder network for seismic data deconvolution, which compresses and reconstructs seismic data using a quantized subset of samples, effectively removing noise and reconstructing seismic waves.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If sparse sampling is used to reduce acquisition costs, then exploration budget is reduced, but data quality and signal amplitude decrease
Solution Approach 1:
The patent replaces conventional mechanical/mathematical deconvolution systems with an LSTM neural network-based autoencoder system. This substitution enables the network to learn complex non-linear relationships in seismic data, achieving superior reconstruction quality from sparse samples without being constrained by traditional filter length limitations
Solution Approach 2:
The patent changes the fundamental parameters of the deconvolution approach by using a data-driven neural network model instead of fixed mathematical transforms. The LSTM network adapts its parameters during training to optimize reconstruction performance, allowing it to handle sparse sampling effectively while maintaining high data quality
2Adaptability or versatility
If irregular sampling is used to adapt to acquisition conditions, then acquisition flexibility is improved, but aliasing is created affecting data processing
Solution Approach 1:
The patent replaces conventional aliasing correction methods based on mathematical transforms with an LSTM neural network that learns to directly reconstruct seismic signals from irregularly sampled data. The network's ability to model temporal dependencies allows it to eliminate aliasing effects without requiring regular sampling patterns
3Productivity
If conventional deconvolution filters are used to reduce computational cost, then processing speed is improved, but filter length is limited reducing performance
Solution Approach 1:
The patent transforms the static deconvolution filter into a dynamic neural network model. The LSTM autoencoder adapts its internal parameters during training to optimize for each specific dataset, enabling both high processing speed through efficient forward propagation and high accuracy through learned optimal filter characteristics that are not constrained by fixed length limitations
4Measurement precision
If LSTM autoencoder is used to improve deconvolution accuracy, then reconstruction quality is improved, but computational complexity increases
Solution Approach 1:
The patent performs computationally intensive training of the LSTM autoencoder in advance during an offline phase. Once trained, the model can perform rapid deconvolution during online processing. This preliminary action separates the heavy computational burden from the actual data processing workflow, achieving both high accuracy and efficient processing
Solution Approach 2:
The patent creates a trained neural network model that serves as a reusable copy of the learned deconvolution process. Once trained on training data, the same model can be applied to multiple different seismic datasets without retraining, significantly reducing computational complexity for subsequent processing tasks
Data Source
AI summary
A system for performing a seismic survey of a geological formation is described. The system includes a seismic transducer, an array of seismic sensors, a transmitter, a receiver, and a computing device. The seismic transducer generates a seismic source wavelet within a geological formation. The array of seismic sensors generates seismic traces. The transmitter transmits the seismic traces. The receiver receives the seismic traces and generates a set of seismic time-series traces. A quantizer compresses the set of seismic time series traces by randomly selecting a subset of samples of the set of seismic time-series traces. A trained long short-term memory (LSTM) neural network based autoencoder deconvolves the randomly selected subset of samples and generates a deconvolved randomly selected subset of samples. The computing device reconstructs the seismic waves reflected from the geological formation from the deconvolved randomly selected subset of samples.