Text-Informed Training Data Filtering for Enhanced Virtual Flow Meter Performance

TL;DR

A new method uses large language models to combine unstructured operational text with sensor data for improved anomaly detection in industrial datasets. This multimodal filtering enhances the quality of training data for virtual flow meters, leading to more accurate predictions than traditional rule-based approaches. The approach reduces model error by leveraging contextual insights from engineer comments and logs.

Why It Matters

Industrial machine learning models often fail due to poor data quality, and manual curation is time-consuming. This method automates high-quality data filtering by using LLMs to interpret operational context, enabling more reliable soft-sensor models in energy and process industries. It offers a scalable solution that improves model performance without requiring extensive domain-specific rules.

Key Findings

• Integrating textual operational logs with sensor data via LLMs improves anomaly detection beyond sensor-only methods.
• The multimodal filtering approach reduces downstream model MAPE from 7.1% to 5.4% on a synthetic oil and gas dataset.
• Engineer comments contain valuable signals for data filtering, demonstrating that unstructured text enhances data quality assessment.