A deep sequence multi-distribution adversarial model for bearing abnormal condition detection. (September 2021)

• Record Type: Journal Article
• Authors: Ou, Xuelian; Wen, Guangrui; Huang, Xin; Su, Yu; Chen, Xuefeng; Lin, Hailong
• Is Part Of: Measurement. Volume 182(2021)
• Publication Date: 2021-09
• Subjects: Condition monitoring -- Anomaly detection -- Sequence data -- Generative adversarial network -- Long short-term memory; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.109529
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: Proposed model that use normal data training are more suitable for practical applications. Apply adversarial learning and LSTM structure to anomaly detection. Multi-distribution features are utilized to construct an anomaly index. DSMDA can prove sufficient robustness for anomalies detection of low SNR signals. Abstract: Time series anomaly detection is one of the key challenges in the field of condition monitoring. Many anomaly detection methods are inefficient and easy to lose effective information due to manual features extracting. Deep learning-based methods can solve the problem effectively, but the detection accuracy is still not satisfactory. In addition, most of the methods cannot take the time-ordered specialty into account which is significant for time-series-based anomaly detection. To address these issues, a novel method named deep sequence multi-distribution adversarial model (DSMDA) is proposed to improve the accuracy of anomaly detection in bearing condition monitoring. The proposed model utilizes the data reconstruction capability of the Variational Autoencoder (VAE) under the framework of generative adversarial network (GAN) to make full use of information. The feedforward neural network layer of VAE is replaced by the long-term and short-term memory (LSTM) layer, which uses the forgetting mechanism of LSTM to effectively avoid the false alarms caused by the excessive influence of the old sequences. Additionally, the fault-attention abnormal state index can be constructed by the real-time spatial distribution and latent spatial distribution features learned by the double discriminators. To verify the effectiveness of the proposed approach, experiments on two public datasets are carried out with only healthy data in training stage that is more suitable for practical industrial applications. The results show that the proposed method is superior to GANomaly and other advanced methods. Furthermore, the 2-D visualization results can indicate the level of fault while the last feature space of the two discriminators is combined and embedded into the visualization, and the fault-attention abnormal state indictor constructed on these features can indicate abnormalities well.