Speaker
Description
Anomaly detection in industrial time-series data is challenged by the scarcity of labeled fault data, limiting the applicability of supervised machine learning methods. This study investigates two complementary approaches for anomaly detection in the Petrobras 3W oil production dataset. A semi-supervised Temporal Convolutional Network (TCN) autoencoder was trained exclusively on normal operating data to identify deviations from learned patterns. In parallel, a supervised XGBoost classifier was developed using statistical features extracted from time-series windows to provide a performance benchmark. The models were evaluated on multiple operational anomaly types, including spurious DHSV closure, severe slugging, and flow instability.
The supervised XGBoost model achieved higher classification performance, highlighting the advantage of access to labeled anomaly examples while also demonstrating the challenges associated with deploying fully supervised methods in data-scarce environments.
| Topic Area | Smart Operations and Maintenance |
|---|