Tensor-DAAD: When tensor meets online early fault detection with transfer learning

Abstract

This paper tries to address the concerns about online early fault detection in non-stop scenarios: (1) no information of running status and operating condition is available for online data; (2) False alarms are raised by irregular noise. This paper proposes a new self-supervised deep tensor domain-adversarial anomaly detection model, named Tensor-DAAD. A contrastive learning network is first built to generate pre-trained features and pseudo-labels for online data. Second, a new tensor hypersphere-adversarial network is designed. By calculating core tensors of two domains’ features, this network builds hypersphere-based one-class detection rules from noisy data, and then realizes rule adaptation between the domains through adversarial training. Finally, a training algorithm with alternating minimization scheme is proposed to seek the optimal tensor decomposition and domain-invariant feature representation. Results on the two widely-used bearing datasets demonstrate that Tensor-DAAD can find early fault occurrence in an earlier location with a much lower false alarm rate.