Unsupervised transfer learning for anomaly detection: Application to complementary operating condition transfer

Anomaly Detectors are trained on healthy operating condition data and raise an alarm when the measured samples deviate from training distribution. This means that used to train model should be sufficient in quantity representative of conditions. But for industrial systems subject changing conditions, acquiring such comprehensive sets requires a long collection period delay point at which anomaly detector can put operation. A solution this problem is perform unsupervised transfer learning (UTL), complementary between different units. In literature however, UTL aims finding common structure datasets, clustering or dimensionality reduction. Yet, task transferring combining has not been studied. Our proposed framework designed conditions units completely way more robust detectors. It differs, thereby, other works as it focuses one-class classification problem. The methodology enables detect anomalies only experienced by end-to-end uses adversarial deep ensure alignment units' distributions. introduces new loss, inspired reduction tool, enforce conservation inherent variability each dataset, state-of-the art once-class approach anomalies. We demonstrate benefit using three open source datasets.