Anomaly Detection for Industrial Fields Using Grounded Labels.

Anomaly detection in real-world industrial settings, such as oil and gas fields, presents unique challenges due to the lack of pre-labeled fault data. Unlike test rigs where faults are intentionally induced and recorded, operational fields provide raw sensor histories without corresponding fault logs, making it difficult to train and validate anomaly detection systems. This research addresses this gap by working with the open Volve field data from Equinor. A key innovation is the creation of "grounded" anomaly labels, which are not merely statistical patterns but are cross-referenced with the field's engineering documentation to ensure physical relevance. The methodology behind each label is also publicly released for transparency. The study evaluated these constructed labels using both an unsupervised baseline and a compact dual-head supervised model. The unsupervised detector successfully identified regions consistent with the rule-based labels, confirming their non-arbitrary nature. The supervised model, adapted from prior work on defect detection, proved effective in recovering event presence and type across unseen wells, though its temporal localization of events was approximate. The dataset, labels, code, and models are all publicly available.