Synthetic to Real Framework based on Convolutional Multi-Head Attention and Hybrid Domain Alignment

Abstract

Domain adaptation (DA) has obtained remarkable results in unsupervised fault diagnosis methods. However, its performance is highly related to two significant factors: Firstly, proposed DA methods should alleviate the global and local distribution gap to precisely matching all distributions in the source and target domain. Most distance-based DA methods assume global DA or concentrate only on local aligning distributions on the country. Secondly, the generalization of most proposed unsupervised fault diagnosis methods relies on labeled faulty data collected from sensors. Contrarily, collected data in real-world scenarios are mostly unlabeled, which considerably declines the model's generalization. We proposed an unsupervised synthetic to the real framework to overcome two significant challenges. A convolution multi-head attention network based on hybrid multi-layer domain alignment (CMHA-HMLDA) is conducted to align global and local distributions. It also alleviates the gap between real and synthetic data more accurately to maintain a robust data-driven model for bearing fault diagnosis. Furthermore, the proposed method is reliable in real scenarios because of transferring employed knowledge of labeled synthetic data into unlabeled real data. To show the supervisory of our proposed method in diagnosing unlabeled real health states, we validated it with a synthetic dataset made from the benchmark bearing Case Western Reserve University (CWRU) dataset. We compared it with recently published unsupervised fault diagnosis methods. Consequently, we achieved state-of-art results that show our proposed method is capable of realizing unlabeled real bearing faults from synthetic data, and it is practical in real-world scenarios.