Do Large Foundation Models Improve Time Series Segmentation? An Industrial Case Study in Oil and Gas Drilling

Imane Khaouja, Amine EL KHAIR, Abdallah Benzine, Sebastiaan Buiting, Soumyadipta Sengupta, Youssef Tamaazousti

Keywords: Time Series Segmentation, Foundation Models, Machine Learning, Oil and Gas, Drilling Operations, Chronos-Bolt, Moment, Moirai, Timer, TimeMOE, GPT4TS

Abstract:

Segmenting time series into meaningful events is critical in domains like drilling, where accurate activity recognition enables operational optimization and real-time decision-making. Yet, segmentation remains challenging due to noise and multivariate complexity. Recently, Foundation Models for Time Series (FM4TS) have emerged as general-purpose solutions, but their effectiveness for segmentation is unclear.

In this study, we benchmark popular FM4TS (both pretrained and trained from scratch) against a fully convolutional network (FCNN) baseline on two tasks: a simple univariate and a complex multivariate segmentation problem. We also assess how performance scales with data size.

Results show CNNs are strong baselines, often outperforming or matching FM4TS. Pretraining offers limited or even negative impact on FM4TS segmentation performance, highlighting challenges in transferring segment-level features. Interestingly FM4TS seems to scale better with more data, suggesting potential advantages in data-rich settings.

Submission Number: 35