Physics-Informed Neural Network for Sensor Data Denoising
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for denoising sensor data in wellbore operations, such as autoencoders, are inadequate in distinguishing between relevant physical phenomena and system noise, leading to inaccuracies due to factors like acoustic vibrations and electrical discharges.
Innovation Solution
A physics-informed deep neural network (PDNN) model is used to generate reduced-noise sensor data, which is then utilized to train an autoencoder to effectively denoise raw sensor data and detect outliers by incorporating a physics-based cost function that differentiates between relevant physical processes and system noise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If an autoencoder is used to denoise sensor data, then the data can be processed and reduced in noise, but the autoencoder cannot distinguish between relevant physical phenomena and system noise, leading to inaccuracies
Solution Approach 1:
The patent introduces a physics-informed layer as an intermediary component between the input and the autoencoder. This layer incorporates domain knowledge about physical phenomena to guide the denoising process, enabling the system to distinguish between relevant physical signals and system noise while maintaining measurement precision.
Solution Approach 2:
The patent modifies the cost function parameters of the autoencoder by incorporating physics-based constraints and domain knowledge. This changes the optimization landscape to favor solutions that preserve physically meaningful patterns while removing noise, thereby improving denoising accuracy without losing information about relevant physical phenomena.
2Measurement precision
If a physics-informed deep neural network is used to denoise sensor data, then the accuracy and efficiency of denoising is significantly improved, but the device complexity increases
Solution Approach 1:
The patent divides the denoising system into distinct modular components: a physics-informed deep neural network layer and an autoencoder. This segmentation allows each component to have a specialized function, improving overall accuracy while making the complex system more manageable and interpretable.
Solution Approach 2:
The physics-informed layer performs preliminary processing of sensor data before it enters the autoencoder. By pre-processing the data with domain knowledge embedded in the neural network, the system reduces the burden on the autoencoder and achieves better denoising accuracy without proportionally increasing overall complexity.
Data Source
AI summary
A physics-influenced deep neural network (PDNN) model, or a deep neural network incorporating a physics-based cost function, can be used to efficiently denoise sensor data. To generate the PDNN model, noisy sensor data is used as training data input to a deep neural network and training output is valuated with a cost function that incorporates a physics-based model. An autoencoder can be coupled to the PDNN model and trained with the reduced-noise sensor data which is output from the PDNN during training of the PDNN or with a separate set of sensor data. The autoencoder detects outliers based on the reconstructed reduced-noise sensor data which it generates. Denoising sensor data by leveraging an autoencoder which is influenced by the physics of the underlying domain based on the incorporation of the physics-based model in the PDNN facilitates accurate and efficient denoising of sensor data and detection of outliers.