Are Time Series Foundation Models Ready for Oil and Gas Drilling Anomaly Detection?

Anomaly detection of multivariate time-series data is an important component of oil and gas drilling anomaly detection, including early identification of hookload and torque events associated with stuck-pipe risk. Although recent Time Series Foundation Models (TSFM) have shown strong forecasting performance, their value for industrial anomaly detection tasks remains unclear. In this paper, we evaluate two forecasting-based TSFM models, Chronos-2 and Toto, against lightweight CNN, CNN-FNO, and MLP baselines for hookload and torque anomaly detection. Our experiments show that Chronos-2 obtains the highest F1 score on both tasks. However, the gain is task-dependent. Domain specific forecasting-based finetuning gives only a small improvement on hookload, while it improves Chronos-2 substantially on torque anomaly detection. We also find that a 5M-parameter CNN-FNO reaches near-parity with the 120M-parameter Chronos-2 on the torque task, with much lower inference time, but does not close the gap on hookload. These results suggest that newer TSFM models can be useful for drilling anomaly detection, particularly when combined with domain-specific forecasting based finetuning. At the same time, their advantage is not uniform across anomaly types, and simpler architectures remain strong candidates for real-time deployment. The study highlights the need to evaluate TSFM not only by accuracy, but also by latency, and cost.