Optical Sensor Data Transformation Neural Networks for Downhole Fluid Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Direct modeling to determine multivariate correlations between optical sensor responses and diverse fluid compositions in hydrocarbon wellbores is cost-prohibitive due to the need for sensor-based calibration, limiting the effectiveness of downhole fluid analysis.
Innovation Solution
A concatenated optical computing neural network (COCN) is employed, combining in-field sensor measurements and simulation data, utilizing a progressive modeling scheme with optical data transformation neural networks to enhance fluid characterization, including a multi-layer perceptron architecture and Principle Component Analysis, to reduce dimensionality and improve data mapping.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If direct modeling with sensor-based calibration is used to determine multivariate correlation between optical sensor responses and fluid compositions, then measurement precision is improved, but cost increases to prohibitive levels
Solution Approach 1:
The patent applies preliminary action by pre-calibrating optical data transformation neural networks using simulated optical sensor responses generated from downhole fluid simulation data before actual downhole deployment. This pre-calibration establishes initial multivariate correlations between synthetic optical responses and fluid compositions, enabling the system to achieve accurate fluid characterization without requiring expensive sensor-based calibration during field operations. The pre-trained networks are then deployed downhole where they process actual optical sensor measurements with preserved accuracy.
2Measurement precision
If progressive modeling with concatenated neural networks is used to integrate simulated and measured data, then fluid characterization accuracy is enhanced, but device complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the fluid characterization system into two distinct neural network components: an optical data transformation neural network that processes optical sensor measurements and a fluid characterization neural network that determines fluid properties. These segmented networks are concatenated in sequence, where the output of the first network serves as input to the second. This segmentation allows each network to specialize in specific transformations while maintaining overall system accuracy and enabling independent training and optimization of each component.
Solution Approach 2:
The patent applies the nested doll principle by nesting the optical data transformation neural network within the overall fluid characterization system. The first network transforms optical sensor responses into intermediate representations, which are then fed into the second network for fluid property prediction. This nested structure allows the system to process complex multivariate relationships through multiple layers of transformation, with each network layer building upon the previous one to achieve enhanced characterization accuracy.
Data Source
AI summary
Disclosed herein are examples embodiments of a progressive modeling scheme to enhance optical sensor transformation networks using both in-field sensor measurements and simulation data. In one aspect, a method includes receiving optical sensor measurements generated by one or more downhole optical sensors in a wellbore; determining synthetic data for fluid characterization using an adaptive model and the optical sensor measurements; and applying the synthetic data to determine one or more physical properties of a fluid in the wellbore for which the optical sensor measurements are received.