Semi-supervised double attention guided assessment approach for remaining useful life of rotating machinery

Many data-driven methods attempt to derive key degradation features from vibration data. However, information consistency and local semantics are not considered when minimizing feature distribution discrepancy in global adaptation learning, resulting in limitations. Meanwhile, the traditional convolution-based feature extractor is difficult to establish the global connection between features. To address the above issues, a semi-supervised double attention guided assessment approach (SDAGA) is proposed to evaluate the remaining useful life of rotating machinery. Specifically, a multi-level deformable convolution module is designed to establish global connections between features and further extract global information. Also, information consistency and multiscale semantics can be maintained by U-shaped sampling and double attention module in a semi-supervised domain adaptation, making it learn more robust invariant degradation features. Extensive experiments are conducted on public and experimental datasets to validate the effectiveness and superiority of SDAGA compared to state-of-the-art methods.