Qwen3.5-powered agent explains oil well anomalies with 89.7% detection accuracy

A team from Brazil's academic and research network has developed an explainable LLM agent layer that sits downstream of existing open-world learning (OWL) pipelines for oil well anomaly detection. The agent, evaluated on 989 real well-file segments from the public 3W dataset, pairs with upstream methods that use autoencoder-based detection and Mahalanobis-based novelty detection. It runs on Qwen3.5-397B-A17B, a Mixture-of-Experts model served through NVIDIA NIM, and converts structured sensor metrics and upstream assertions into natural-language justifications, confidence-ranked critiques, and consolidated labels for novel anomaly clusters.

The agent is not a standalone classifier; it serves as an auditable companion that confirms upstream decisions when sensor evidence is consistent, justifies those decisions in operator-readable language, flags implausible labels, and names new types of anomalies so engineers receive consolidated, human-readable labels. Across three studies it achieved 35.1% top-1 / 63.9% top-3 accuracy on nine classes, 71.7% top-2 validation with precision 0.91 across seven probed classes, and 89.7% novelty detection with stable naming on five of seven hidden classes. The goal is to close the explainability gap that currently prevents OWL pipelines from being deployed in operational oilfield settings.