A novel multiscale feature adversarial fusion network for unsupervised cross-domain fault diagnosis. (15th August 2022)

• Record Type: Journal Article
• Authors: Shi, Yaowei; Deng, Aidong; Deng, Minqiang; Xu, Meng; Liu, Yang; Ding, Xue
• Is Part Of: Measurement. Volume 200(2022)
• Publication Date: 2022-08-15
• Subjects: Fault transfer diagnosis -- Rotating machinery -- Domain adaptation -- Adversarial training -- Multiscale feature; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2022.111616
• Languages: English
• ISSNs: 0263-2241

Abstract

Graphical abstract: Highlights: A MFAFN is proposed for unsupervised cross-domain fault diagnosis. The TDAM is designed to reinforce the feature and assist in network training. The DAA strategy ensures the effectiveness of domain adaptation. The MFAFN has powerful feature extraction and domain adaptation abilities. Abstract: Behind the brilliance of traditional deep learning-based diagnosis methods, the assumption that training data and test data share the same distribution greatly hinders their further application. Because the data distribution discrepancy is common and inevitable in real industrial scenarios due to operating condition variation, it will significantly degrade models' diagnosis performance. Moreover, scarce labeled data can be obtained, and labeling sufficient data is extremely difficult and expensive in engineering applications. Considering these challenges, this paper proposes a novel multiscale feature adversarial fusion network (MFAFN) for rotating machinery fault transfer diagnosis. In our method, the multiscale structural network is employed to extract abundant and complementary multiscale features. The key highlight of MFAFN is that the transferability-based duplex attention mechanism (TDAM) is elaborated and bidirectionally coupled into the network training. Benefiting from TDAM, the representation and learning of the extracted features at different scales are differentially enhanced, thus improving the fused shared feature's transferability and model's adaptability. Furthermore, a double-level adversarial training strategy is implemented to ensure effective adaptation. Therefore, MFAFN can learn domain-invariant diagnosis knowledge rich in discriminative fault information, thereby performing better on the unlabeled target domain. Experimental results of extensive diagnosis tasks built on two datasets and comparisons with other methods validate MFAFN's effectiveness and superiority.