Paper review · October 16, 2026
Improving Virtual Flow Meter Performance with Training Data Filtering Using Text Information
Text-Informed Training Data Filtering for Enhanced Virtual Flow Meter Performance
An LLM looks at sensor data together with operator comments to filter out anomalous data, reducing the virtual flow meter's prediction error compared with sensor-only filtering.
Problem
Industrial sensor data contains many anomalies, such as sensor failures, process upsets and input errors, which reduce model accuracy. Rule-based filters catch only statistical outliers. They miss data that is within the normal range but wrong in context, that is, anomalies that can only be identified by reading operator comments or logs. Until now, experts have cleaned the data by manually cross-checking the two.
Method
They used the ManyWells multiphase flow simulator to create a synthetic dataset of 13 wells. They injected six types of anomalies, such as scaling, hydrates and wax, and attached operator comments in a mix of Norwegian and English. 44.6% of anomalous data points have no comment. An LLM (Claude Sonnet 4.6) was given a task description, a variable dictionary, three reference wells with labels revealed (few-shot examples) and the target well's data in JSON, and asked to output include/exclude (0/1) for each data point. After filtering the data with this mask, a simple physics-based soft sensor was trained and evaluated with MAPE, MSE and W10. Baselines were Tukey fences, robust Mahalanobis distance, no filter and a ground-truth filter (Oracle).
Results
- In the ablation, multimodal LLM filtering given both sensors and comments lowered downstream model MAPE to 5.4%, compared with 7.1% for LLM filtering given sensors only.
- The authors report that multimodal LLM filtering produced better soft sensor performance than rule-based filters such as Tukey and Mahalanobis.
- The authors conclude that operator comments contain filtering signals not visible from sensor data alone.
Limitations
The dataset is synthetic rather than real operational data, and it is small: 311 training data points from 10 wells. It was simplified so that anomalous data never carries a 'normal' comment, so comments may be cleaner than in practice. The downstream model is a simple steady-state physics model, and only one LLM was run, 3 times. Effects such as reservoir changes over time were not modeled. The provided text was cut off, so exact numerical comparisons with rule-based filters could not be checked.
Why it matters
It shows that LLMs could automate the work of cleaning industrial data, where experts used to cross-check dashboards and logs by hand. It is a data-centric AI example that uses text logs for data quality control.
Key terms
- Virtual Flow Meter
- A soft sensor model that estimates the flow rate of a well, which is hard to measure directly, from other measurements such as pressure.
- Contextual Anomaly
- Data whose values fall within a statistically normal range but cannot be trusted once the circumstances at the time, such as an equipment failure, are taken into account.
- Tukey fences
- A classic rule-based method that flags values lying more than 1.5 times the interquartile range (IQR) outside the quartiles as outliers.
- MAPE
- Mean absolute percentage error. It shows by what percentage, on average, predictions differ from actual values.
Original
Authors: Kristoffer Hjelmen Nesland, Bjarne Grimstad, Antorweep Chakravorty, Dan Sui · Posted 2026-10-14 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.