Modified DSAN for unsupervised cross-domain fault diagnosis of bearing under speed fluctuation

Existing researches about unsupervised cross-domain bearing fault diagnosis mostly consider global alignment of feature distributions in various domains, and focus on relatively ideal diagnosis scenario under the steady speeds. Therefore, unsupervised feature adaptation between all the corresponding subdomains under speed fluctuation remains great challenges. This paper proposes a modified deep subdomain adaptation network (MDSAN) for more practical and challenging cross-domain diagnostic scenarios from the fluctuating speeds to steady speeds. Firstly, to extract the representative features and effectively suppress negative transfer, a novel shared feature extraction module guided by multi-headed self-attention mechanism is constructed. Then, a new trade-off factor is designed to improve the convergence performance and optimization process of MDSAN. The proposed method is used for analyzing experimental bearing vibration data, and the results show that it has higher diagnostic accuracy, faster convergence, better distribution alignment, and is more suitable for unsupervised cross-domain fault diagnosis under speed fluctuation scenario compared with the existing methods.