DUnet Architecture for Seismic Data Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Seismic data processing is complex due to noise and interference, making it difficult to integrate deep learning methods effectively into industrial toolkits, particularly in interpreting and quality-controlling operations performed by deep neural networks and determining suitable architectures for specific tasks.
Innovation Solution
The DUnet architecture, which includes a contractive path for data extraction, an expansive path for feature expansion using skip connections, and a model level for producing denoised images, is used to improve seismic data processing by enhancing output quality and interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning methods are used for seismic data processing, then output quality and interpretability are improved, but device complexity and difficulty of integration increase
Solution Approach 1:
The DUnet architecture segments the deep learning model into distinct functional paths: a contractive path for feature extraction and an expansive path for feature reconstruction. This segmentation makes the complex model more manageable and interpretable while maintaining high output quality for seismic data processing.
Solution Approach 2:
The code bottleneck acts as an intermediary between the contractive path and expansive path, facilitating controlled information flow and enabling quality control of the deep learning operations. This intermediary structure helps manage device complexity by providing a clear interface for interpreting model operations.
2Measurement precision
If deep learning methods are used for seismic data processing, then output quality and interpretability are improved, but ease of operation and quality control become more difficult
Solution Approach 1:
The skip connections provide feedback mechanisms that allow intermediate features to be reused and refined throughout the network. This feedback structure improves ease of operation by enabling controlled exploration of feature transformations and facilitating quality control of the processing operations.
Solution Approach 2:
The model level connected to both contractive and expansive paths serves as an intermediary that coordinates the deep learning operations, making the system easier to operate and control while maintaining high output quality.
3Ease of manufacture
If conventional deep neural network architectures are used, then integration is simpler, but output quality and interpretability are reduced
Solution Approach 1:
By segmenting the network into contractive and expansive paths with a code bottleneck, the architecture maintains relative integration simplicity while dramatically improving output quality and interpretability compared to conventional flat DNN architectures.
Solution Approach 2:
The architecture introduces a dimensional transformation by compressing features to a code bottleneck and then expanding them, creating a U-shaped processing flow that enhances both output quality and interpretability while preserving ease of integration through modular design.
Data Source
AI summary
A DUnet engine produces a processed image of seismic data acquired over an underground formation. The DUnet engine includes: a contractive path that performs multilayer convolutions and contraction to extract a code from the seismic data input to the DUnet, an expansive path configured to perform multilayer convolutions and expansion of the code, using features provided by the contractive path through skip connections, and a model level that performs multilayer convolutions on outputs of the contractive path and expansive paths to produce the processed image and/or an image that is a difference between the processed image and the seismic data. A fraction of the seismic data may be selected for training the DUnet engine using an anchor method that automatically extends an initial seismic data subset, based on similarity measurements. A reweighting layer may further combine inputs received from layers of the DUnet model to preserve signal amplitude trend.