Deep attention SMOTE: Data augmentation with a learnable interpolation factor for imbalanced anomaly detection of gas turbines

• Published in: Computers in industry Vol. 151; p. 103972
• Main Authors: Liu, Dan, Zhong, Shisheng, Lin, Lin, Zhao, Minghang, Fu, Xuyun, Liu, Xueyun
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2023
• Subjects: Anomaly detection; Data augmentation; Gas turbine; Imbalanced data; Inter-class overlap; Inter-class overlap; Imbalanced data; Data augmentation; Gas turbine; Anomaly detection
• ISSN: 0166-3615; 1872-6194
• DOI: 10.1016/j.compind.2023.103972

Anomaly detection of gas turbines faces the significant challenges of data imbalance and inter-class overlap. In this paper, we develop a novel data augmentation method, namely deep attention synthetic minority over-sampling technique with the Encoder-Decoder (DA-SMOTE-ED), which serves as a key step in our hybrid re-sampling scheme. To reduce the risk of generating noise data, on one hand, the DA-SMOTE-ED leverages an Encoder-Decoder to learn a class-separable feature space to weaken the effect of inter-class overlap. On the other hand, an attention module is applied to assign proper interpolation factors to generate synthetic samples that stay off the aggregation area of normal samples. Moreover, synthetic samples are generated in the learnable feature space, mapped back to the original space, and merged with under-sampled samples to form the balanced dataset. Finally, the superiority of the developed method is validated through two case studies including the real monitoring data of gas turbines and the modified version of the commercial modular aero-propulsion system simulation (C-MAPPS) dataset. More specifically, its average balanced accuracy is 91.77 % on the gas turbine dataset, yielding 3.67 %, 6.4 %, and 5.56 % improvements compared to the SMOTE-ENN, TimeGAN, and AugmentTS, respectively. •A new hybrid re-sampling scheme is proposed to overcome the class imbalance issue.•A feature space with inter-class separability is learned for data augmentation.•An attention module is designed to learn adaptive interpolation factors.•The adaptive factors generate high-quality synthetic samples in the feature space.•The scheme is validated using the gas turbine dataset and the public dataset.