Deep Segmental Denoising Neural Network for Seismic Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional noise reduction techniques for seismic data, such as time-frequency based denoising methods, face challenges in effectively separating signal from noise when noise occupies the same frequency band as the seismic signal, often requiring optimal thresholding functions that are difficult to determine, leading to either signal attenuation or ineffective noise suppression.
Innovation Solution
A deep segmental denoising neural network is employed, which transforms recorded seismic signals into a time-frequency representation using discrete cosine transforms and utilizes a deep convolutional neural network trained with segments of noisy and clean spectra to learn a mapping function for denoising, effectively enhancing the signal-to-noise ratio without relying on thresholding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If conventional bandpass or spectral filtering techniques are used to remove noise, then random wide-band noise can be effectively removed, but noise occupying the same frequency band as the seismic signal cannot be suppressed
Solution Approach 1:
The patent transforms the seismic signal from the time domain to the time-frequency domain using Short-Time Fourier Transform (STFT), changing the representation parameters to enable frequency-based noise separation. This parameter transformation allows the neural network to operate on spectral features rather than raw time-domain signals, effectively addressing noise in the same frequency band by learning frequency-domain patterns.
Solution Approach 2:
The patent introduces a deep neural network as an intermediary between the noisy time-frequency representation and the denoised output. The CNN-based denoising network acts as a mediator that learns to separate signal from noise in the time-frequency domain, then reconstructs the clean signal through inverse transform, bridging the gap between conventional filtering limitations and effective denoising.
2Object-affected harmful factors
If spatial prediction filters are used to suppress noise, then noise in the same frequency band can be reduced, but the seismic signal is substantially altered and subsequent analysis is degraded
Solution Approach 1:
The patent replaces conventional mechanical signal processing operations (spatial prediction filters with fixed parameters) with a data-driven neural network system. The CNN learns optimal filtering operations from training data, substituting rigid mathematical filters with adaptive learned filters that preserve signal characteristics while removing noise, thereby maintaining signal fidelity.
Solution Approach 2:
The patent performs preliminary transformation of the seismic signal into the time-frequency domain before denoising operations. By pre-processing the signal with STFT and extracting spectral features, the system prepares the data in a form that enables more effective and signal-preserving denoising, avoiding direct manipulation of the raw time-domain signal that would degrade fidelity.
3Object-affected harmful factors
If time-frequency transformation and thresholding techniques are used for denoising, then noise can be attenuated, but optimal threshold function is difficult to determine and signal attenuation occurs
Solution Approach 1:
The patent employs a deep neural network that automatically learns the optimal denoising strategy from training data without requiring manual threshold function design. The CNN self-adjusts its filtering parameters and decision boundaries during training, replacing the complex manual optimization of threshold functions with an automated learning process that adapts to the specific characteristics of the seismic data.
Solution Approach 2:
The patent changes from fixed thresholding parameters to learned parameters through neural network training. Instead of using predetermined threshold functions that require complex optimization, the system transforms the parameters into learnable weights and biases in the CNN, which are automatically optimized during training to achieve effective denoising without signal attenuation.
4Object-affected harmful factors
If transformation-based denoising methods are used to suppress noise in overlapping frequency bands, then sparsity is promoted for signal representation, but the methods are not very effective in practice
Solution Approach 1:
The patent combines multiple processing stages into a composite denoising system: time-frequency transformation (STFT), deep neural network processing (CNN with multiple convolutional layers), and inverse transformation. This composite approach integrates the advantages of frequency-domain analysis with the power of deep learning, creating a more reliable denoising system than single-method approaches, effectively handling noise in overlapping frequency bands.
Data Source
AI summary
An apparatus, computer readable storage medium, and method for deep segmental denoising neural network for microseismic data is described. The apparatus includes a seismic data recording network with geophones each having a seismic data receiver and configured to record microseismic waves as a seismic trace received from a geological formation, a preprocessing stage and a deep neural network. The preprocessing stage transforms the recorded signal trace to a time-frequency representation as real number values. The deep neural network generates a denoised signal from the time-frequency representation. The deep neural network is trained based on a segment of noisy spectra and a clean spectra segment to learn a mapping function that generates the segment of the denoised microseismic signal.