Towards Building a Distributed Virtual Flow Meter via Compressed Continual Learning.

Area of Science:

• Machine Learning
• Fluid Dynamics
• Data Science

Background:

• Accurate fluid flow estimation is crucial for distributed virtual flow meters.
• Current machine learning algorithms struggle with predictions outside training data distribution.
• Continual learning (CL) offers a potential solution but faces catastrophic forgetting.

Purpose of the Study:

• To explore the continual learning (CL) paradigm for robust fluid flow estimation in pipelines.
• To address the challenge of catastrophic forgetting in CL.
• To improve the accuracy of machine learning models for predicting fluid flow characteristics.

Main Methods:

• Proposed a novel approach to mitigate catastrophic forgetting in CL.
• Utilized compressive learning to compress distributed sensor data.
• Increased the capacity of the CL memory bank for enhanced learning.

Main Results:

• Achieved approximately 8% accuracy improvement on a real-world oilfield dataset compared to other CL algorithms.
• Demonstrated noticeable accuracy improvements on CL benchmark datasets.
• Reached state-of-the-art accuracies of 80.83% and 88.91% on CIFAR-10 (blurry10 and blurry30 settings).

Conclusions:

• The proposed compressive learning approach effectively addresses catastrophic forgetting in CL.
• This method significantly enhances the accuracy of fluid flow estimation in pipelines.
• The approach shows broad applicability, achieving state-of-the-art results on benchmark image datasets.