IEEE Internet of Things Journal, Год журнала: 2024, Номер 11(12), С. 21895 - 21903

Опубликована: Март 18, 2024

The performance of neural networks is directly affected by the features obtained from backbones fault diagnosis networks. To obtain clear and improve networks, this paper constructs a new block based on linear transformation. Firstly, feature vector divided into decisive component an invalid component. Then, it worth noting that orthogonality these two components beneficial to model learning. According this, are extracted using spaces constructed relationships between four fundamental sub-spaces matrix. In sub-spaces, row space null employed extract useless component, respectively. Both implemented layers designed as encoder-decoder structure ensure existence space. spaces, constraint term proposed modify their weights. Lastly, cosine similarity input entirely. When incorporating some classic classifying they can achieve improved accuracy. Moreover, when comparing conventional spatial attention mechanisms, module demonstrates superior overall performance, including accuracy, antinoise ability, generalization ability.

Язык: Английский