Deep domain-adversarial anomaly detection with partial gradient reversal

Abstract

In industrial settings, outliers can significantly affect equipment stability and potentially jeopardize system’s regular operation and performance. Especially when confronted with new equipment or operating conditions, it becomes imperative to establish a precise model with a limited dataset to swiftly and accurately detect and manage outliers. To tackle this issue, we introduce a transfer learning-based approach aimed at quickly adapting to novel environments and constructing efficient anomaly detection models. This strategy merges transfer learning with conventional anomaly detection techniques to create resilient models within a weakly supervised framework. In contrast to conventional methods disregarding unknown outliers, our approach incorporates a Gradient Partial Reversal technique, employing a domain adversarial mechanism to gently segregate outliers at a distinct level from the anomaly detection algorithms. This strategy yields training outcomes comparable to supervised models. To validate the efficacy of our model, we conducted experiments across three scenarios: digit image detection (utilizing the MNIST and USPS datasets), object recognition (employing the Office–Home dataset), and rolling bearing anomaly detection. Our results show that the proposed algorithm significantly outperforms existing state-of-the-art methods in terms of detection accuracy and robustness.