Feature Enhancement via Linear Transformation and Its Application in Fault Diagnosis

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